add her demo

fixbug
2024-09-07 20:13:14 +08:00 · 2024-09-07 17:15:28 +08:00 · 2024-09-07 16:36:13 +08:00 · 2024-09-07 16:13:24 +08:00 · 2024-09-07 15:14:07 +08:00 · 2024-09-07 10:29:15 +08:00
173 changed files with 185011 additions and 5530 deletions
@@ -0,0 +1,21 @@
+name: Publish to Comfy registry
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - main
+    paths:
+      - "pyproject.toml"
+
+jobs:
+  publish-node:
+    name: Publish Custom Node to registry
+    runs-on: ubuntu-latest
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v4
+      - name: Publish Custom Node
+        uses: Comfy-Org/publish-node-action@main
+        with:
+          ## Add your own personal access token to your Github Repository secrets and reference it here.
+          personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}
@@ -1,4 +1,8 @@
 __pycache__/
 https/
 nodes/config.json
-workflow/my_workflow.json
+workflow/my_workflow.json
+workflow/my_workflow_app.json
+workflow/prompt_result.json
+app/*
+workflow/prompt_result.json
@@ -0,0 +1 @@
+mixlabnodes.com
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 shadow
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
@@ -1,109 +1,339 @@
-## 
-v0.3.0 🚀🚗🚚🏃‍
+![](https://img.shields.io/github/release/shadowcz007/comfyui-mixlab-nodes)

- Added support for setting proxies: HTTP_PROXY, HTTPS_PROXY, http_proxy, https_proxy ✅
+> 适配了最新版 comfyui 的 py3.11 ，torch 2.3.1+cu121
+> [Mixlab nodes discord](https://discord.gg/cXs9vZSqeK)

- Added a new Speech feature node, enabling the use of a voice assistant: SpeechRecognition & SpeechSynthesis 🎙️
+商务合作请联系 389570357@qq.com
+For business cooperation, please contact email 389570357@qq.com

- Added TextImage node, allowing conversion of text into image format 📷
+##### `最新`：

- Added SvgImage node, enabling layout parsing and poster generation in conjunction with the Layer class node 🖼️
+- [增加 Her 的DEMO页面，和数字人对话](https://github.com/shadowcz007/ComfyUI-Backend-MixlabNodes/blob/main/workflow/her_demo_workflow.json)

- Added an experimental 3DImage node for loading 3D models 🌟
+- 右键菜单支持 text-to-text，方便对 prompt 词补全，支持云LLM或者是本地LLM。

+- 增加 MiniCPM-V 2.6 int4
+
+This is the int4 quantized version of MiniCPM-V 2.6.
+Running with int4 version would use lower GPU memory (about 7GB).
+
+- 移动端适配、修改 app 模式的 Mask 编辑器
+
+- 增加 p5.js 作为输入节点
+  [workflow](./workflow/p5workflow.json)
+  [workflow2](./workflow/p5-video-workflow.json)
+
+- App 模式增加 batch prompt，批量提示词，可以把动态提示词批量组成后运行
+
+![alt text](./assets/1722517810720.png)
+
+- 增加 API Key Input 节点，用于管理 LLM 的 Key,同时优化 LLM 相关节点，为后续 agent 模式做准备
+
+- 增加 SiliconflowLLM，可以使用由 Siliconflow 提供的免费 LLM
+
+<!-- - ChatGPT 节点支持 Local LLM（llama.cpp），Phi3、llama3 都可以直接一个节点运行了。模型下载后，放置到 `models/llamafile/` -->
+
+<!--
+强烈推荐：
+[Phi-3-mini-4k-instruct-function-calling-GGUF](https://huggingface.co/nold/Phi-3-mini-4k-instruct-function-calling-GGUF)
+
+[Phi-3-mini-4k-instruct-GGUF](https://huggingface.co/lmstudio-community/Phi-3-mini-4k-instruct-GGUF/tree/main)，备选：[llama3_if_ai_sdpromptmkr_q2k](https://hf-mirror.com/impactframes/llama3_if_ai_sdpromptmkr_q2k/tree/main)
+
+- 右键菜单支持 image-to-text，使用多模态模型，多模态使用 [llava-phi-3-mini-gguf](https://huggingface.co/xtuner/llava-phi-3-mini-gguf/tree/main)，注意需要把llava-phi-3-mini-mmproj-f16.gguf也下载
+
+![](./assets/prompt_ai_setup.png)
+![](./assets/prompt-ai.png) -->
+
+#### `相关插件推荐`
+
+[comfyui-liveportrait](https://github.com/shadowcz007/comfyui-liveportrait)
+
+[Comfyui-ChatTTS](https://github.com/shadowcz007/Comfyui-ChatTTS)
+
+[comfyui-sound-lab](https://github.com/shadowcz007/comfyui-sound-lab)
+
+[comfyui-Image-reward](https://github.com/shadowcz007/comfyui-Image-reward)
+
+[comfyui-ultralytics-yolo](https://github.com/shadowcz007/comfyui-ultralytics-yolo)
+
+[comfyui-moondream](https://github.com/shadowcz007/comfyui-moondream)
+
+<!-- [comfyui-CLIPSeg](https://github.com/shadowcz007/comfyui-CLIPSeg) -->
+
+## 🚀🚗🚚🏃 Workflow-to-APP
+
+- 新增 AppInfo 节点，可以通过简单的配置，把 workflow 转变为一个 Web APP。
+- 支持多个 web app 切换
+- 发布为 app 的 workflow，可以在右键里再次编辑了
+- web app 可以设置分类，在 comfyui 右键菜单可以编辑更新 web app
+- 支持动态提示
+- 支持把输出显示到 comfyui 背景（TouchDesigner 风格）
+- 如果转为 web app 打开是空白的，注意检查下插件目录的名字需要是：comfyui-mixlab-nodes(如果是 zip 包下载会多了个-main 的后缀，需要去掉)
+
+![](./assets/微信图片_20240421205440.png)
+
+- Support multiple web app switching.
+- Add the AppInfo node, which allows you to transform the workflow into a web app by simple configuration.
+- The workflow, which is now released as an app, can also be edited again by right-clicking.
+- The web app can be configured with categories, and the web app can be edited and updated in the right-click menu of ComfyUI.
+
+![](./assets/0-m-app.png)
+
+![](./assets/appinfo-readme.png)
+
+![](./assets/appinfo-2.png)
+
+Example:
+
+- workflow
+  ![APP info](./workflow/appinfo-workflow.svg)
+  [text-to-image](./workflow/Text-to-Image-app.json)
+
+APP-JSON:
+
+- [text-to-image](./example/Text-to-Image_3.json)
+- [image-to-image](./example/Image-to-Image_2.json)
+- text-to-text
+
+> 暂时支持 9 种节点作为界面上的输入节点：Load Image、VHS*LoadVideo、CLIPTextEncode、PromptSlide、TextInput*、Color、FloatSlider、IntNumber、CheckpointLoaderSimple、LoraLoader
+
+> 输出节点：PreviewImage 、SaveImage、ShowTextForGPT、VHS_VideoCombine、PromptImage
+
+> seed 统一输入控件，支持：SamplerCustom、KSampler
+
+> 配套[ps 插件](https://github.com/shadowcz007/comfyui-ps-plugin)
+
+> 如果遇到上传图片不成功，请检查下：局域网或者是云服务，请使用 https，端口 8189 这个服务（ 感谢 @Damien 反馈问题）
+
+> If you encounter difficulties in uploading images, please check the following: for local network or cloud services, please use HTTPS and the service on port 8189. (Thanks to @Damien for reporting the issue.)
+
+## 🏃🚗🚚🚀 Real-time Design
+
+> ScreenShareNode & FloatingVideoNode. Now comfyui supports capturing screen pixel streams from any software and can be used for LCM-Lora integration. Let's get started with implementation and design! 💻🌐

 ![screenshare](./assets/screenshare.png)

-
-### SpeechRecognition & SpeechSynthesis
-![f](./assets/audio-workflow.svg)
-
-
-
-### ScreenShareNode & FloatingVideoNode
-> Now comfyui supports capturing screen pixel streams from any software and can be used for LCM-Lora integration. Let's get started with implementation and design! 💻🌐
-
-> 
-
 https://github.com/shadowcz007/comfyui-mixlab-nodes/assets/12645064/e7e77f90-e43e-410a-ab3a-1952b7b4e7da

-
 <!-- [ScreenShareNode](./workflow/2-screeshare.json) -->
+
 [ScreenShareNode & FloatingVideoNode](./workflow/3-FloatVideo-workflow.json)

 !! Please use the address with HTTPS (https://127.0.0.1).

+### SpeechRecognition & SpeechSynthesis

-### LoadImagesFromLocal
-> Monitor changes to images in a local folder, and trigger real-time execution of workflows, supporting common image formats, especially PSD format, in conjunction with Photoshop. 
+![f](./assets/audio-workflow.svg)

-![watch](./assets/4-loadfromlocal-watcher-workflow.svg)
+[Voice + Real-time Face Swap Workflow](./workflow/语音+实时换脸workflow.json)

-[workflow-4](./workflow/4-loadfromlocal-watcher-workflow.json)
+- Preview Audio

+[text-to-audio](./workflow/text-to-audio-base-workflow.json)

 ### GPT
-> Support for calling multiple GPTs.ChatGPT、ChatGLM3 , Some code provided by rui. If you are using OpenAI's service, fill in https://api.openai.com/v1 . If you are using a local LLM service, fill in http://127.0.0.1:xxxx/v1 .  Azure OpenAI:https://xxxx.openai.azure.com 

+> Support for calling multiple GPTs.Local LLM 、 ChatGPT、ChatGLM3 、ChatGLM4 , Some code provided by rui. If you are using OpenAI's service, fill in https://api.openai.com/v1 . If you are using a local LLM service, fill in http://127.0.0.1:xxxx/v1 . Azure OpenAI:https://xxxx.openai.azure.com

-![gpt-workflow.svg](./assets/gpt-workflow.svg)
+[LLM_base_workflow](./workflow/LLM_base_workflow.json)

-[workflow-5](./workflow/5-gpt-workflow.json)
+- SiliconflowLLM
+- ChatGPTOpenAI

+<!-- 最新：ChatGPT 节点支持 Local LLM（llama.cpp），Phi3、llama3 都可以直接一个节点运行了。
+
+Model download,move to :`models/llamafile/`
+
+强烈推荐：[Phi-3-mini-4k-instruct-GGUF](https://huggingface.co/lmstudio-community/Phi-3-mini-4k-instruct-GGUF/tree/main)
+
+备选：[llama3_if_ai_sdpromptmkr_q2k](https://hf-mirror.com/impactframes/llama3_if_ai_sdpromptmkr_q2k/tree/main)
+
+> 如果碰到安装失败，可以尝试手动安装
+
+```
+../../../python_embeded/python.exe -s -m pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu121
+
+../../../python_embeded/python.exe -s -m pip install llama-cpp-python[server]
+
+```
+
+> [Mac](https://llama-cpp-python.readthedocs.io/en/latest/install/macos/)
+
+```
+pip uninstall llama-cpp-python -y
+CMAKE_ARGS="-DLLAMA_METAL=on" pip install -U llama-cpp-python --no-cache-dir
+pip install 'llama-cpp-python[server]'
+```
+
+```
+pip install llama-cpp-python \
+  --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/metal
+``` -->
+
+## Prompt
+
+> PromptSlide
+> ![](./assets/prompt_weight.png)
+
+<!-- ![](./workflow/promptslide-appinfo-workflow.svg) -->
+
+> randomPrompt
+
+![randomPrompt](./assets/randomPrompt.png)
+
+> ClipInterrogator
+
+[add clip-interrogator](https://github.com/pharmapsychotic/clip-interrogator)
+
+> PromptImage & PromptSimplification,Assist in simplifying prompt words, comparing images and prompt word nodes.
+
+> ChinesePrompt && PromptGenerate，中文 prompt 节点，直接用中文书写你的 prompt
+
+![](./assets/ChinesePrompt_workflow.svg)

 ### Layers
+
 > A new layer class node has been added, allowing you to separate the image into layers. After merging the images, you can input the controlnet for further processing.

+> The composite images node overlays a foreground image onto a background image at specified positions and scales, with optional blending modes and masking capabilities. position : 'overall',"center_center","left_bottom","center_bottom","right_bottom","left_top","center_top","right_top"
+
 ![layers](./assets/layers-workflow.svg)

 ![poster](./assets/poster-workflow.svg)

-## Other Nodes
+### 3D
+
+![](./assets/3d-workflow.png)
+![](./assets/3d_app.png)
+[workflow](./assets/Image-to-3D_1.json)
+
+![](./assets/3dimage.png)
+[workflow](./workflow/3D-workflow.json)
+
+### Image
+
+#### LoadImagesToBatch
+
+> Upload multiple images for batch input into the IP adapter.
+
+#### LoadImagesFromLocal
+
+> Monitor changes to images in a local folder, and trigger real-time execution of workflows, supporting common image formats, especially PSD format, in conjunction with Photoshop.
+
+![watch](./assets/4-loadfromlocal-watcher-workflow.svg)
+
+[workflow-4](./workflow/4-loadfromlocal-watcher-workflow.json)
+
+#### LoadImagesFromURL
+
+> Conveniently load images from a fixed address on the internet to ensure that default images in the workflow can be executed.
+
+#### TextImage
+
+> [下载字体](https://drxie.github.io/OSFCC/)放到 `custom_nodes/comfyui-mixlab-nodes/assets/fonts`
+
+#### MiniCPM-VQA Simple
+
+This is the int4 quantized version of MiniCPM-V 2.6.
+Running with int4 version would use lower GPU memory (about 7GB).
+
+[模型](https://huggingface.co/openbmb/MiniCPM-V-2_6-int4)
+
+![alt text](assets/1724308322276.png)
+
+### Style
+
+> Apply VisualStyle Prompting , Modified from [ComfyUI_VisualStylePrompting](https://github.com/ExponentialML/ComfyUI_VisualStylePrompting)
+
+![](./assets/VisualStylePrompting.png)
+
+> StyleAligned , Modified from [style_aligned_comfy](https://github.com/brianfitzgerald/style_aligned_comfy)
+
+### Utils
+
+> The Color node provides a color picker for easy color selection, the Font node offers built-in font selection for use with TextImage to generate text images, and the DynamicDelayByText node allows delayed execution based on the length of the input text.
+
+- [添加了 DynamicDelayByText 功能，可以根据输入文本的长度进行延迟执行。](./workflow/audio-chatgpt-workflow.json)
+
+- [Added DynamicDelayByText, enabling delayed execution based on input text length.](./workflow/audio-chatgpt-workflow.json)
+
+- [使用 CkptNames 对比不同的模型效果](./workflow/ckpts-image-workflow.json)
+
+- [CkptNames compare the effects of different models.](./workflow/ckpts-image-workflow.json)
+
+### Other Nodes
+
+- 增加 Edit Mask，方便在生成的时候手动绘制 mask [workflow](./workflow/edit-mask-workflow.json)

 ![main](./assets/all-workflow.svg)
 ![main2](./assets/detect-face-all.png)

 [workflow-1](./workflow/1-workflow.json)

-> randomPrompt
-
-![randomPrompt](./assets/randomPrompt.png)
-
 > TransparentImage

 ![TransparentImage](./assets/TransparentImage.png)

-
-> Consistency Decoder
-
-[openai Consistency Decoder]( https://github.com/openai/consistencydecoder)
-
-![Consistency](./assets/consistency.png)
-After downloading the OpenAI VAE model, place it in the "model/vae" directory for use.
-https://openaipublic.azureedge.net/diff-vae/c9cebd3132dd9c42936d803e33424145a748843c8f716c0814838bdc8a2fe7cb/decoder.pt
-
-
 > FeatheredMask、SmoothMask

 Add edges to an image.

 ![FeatheredMask](./assets/FlVou_Y6kaGWYoEj1Tn0aTd4AjMI.jpg)

+> LaMaInpainting（需要手动安装）

+- simple-lama-inpainting 里的 pillow 造成冲突，暂时从依赖里移除，如果有安装 simple-lama-inpainting ，节点会自动添加，没有，则不会自动添加。

-### Improvement 
-An improvement has been made to directly redirect to GitHub to search for missing nodes when loading the graph.
+from [simple-lama-inpainting](https://github.com/enesmsahin/simple-lama-inpainting)
+
+- [问题汇总](https://github.com/shadowcz007/comfyui-mixlab-nodes/issues/294)
+
+> rembgNode
+
+"briarmbg","u2net","u2netp","u2net_human_seg","u2net_cloth_seg","silueta","isnet-general-use","isnet-anime"
+
+**_ briarmbg _** model was developed by BRlA Al and can be used as an open-source model for non-commercial purposes
+
+### Enhancement
+
+- Direct "Help" option accessible through node context menu.
+
+- "Nodes Map" feature added to global context menu.
+
+- An improvement has been made to directly redirect to GitHub to search for missing nodes when loading the graph.
+
+*** If not needed, you can comment out ```app.showMissingNodesError``` in the ```ui_mixlab.js``` file.
+
+![help](./assets/help.png)

 ![node-not-found](./assets/node-not-found.png)


+- Right-click shortcut
+
+右键菜单支持 text-to-text，方便对 prompt 词补全，支持云LLM或者是本地LLM。
+The right-click menu supports text-to-text conversion, facilitating prompt word completion, and supports cloud LLMs or local LLMs.
+
+Local LLM API example:```http://localhost:1234/v1```
+
+![alt text](./assets/1724380841822.png)
+
+
 ### Models
-[Download CLIPSeg](https://huggingface.co/CIDAS/clipseg-rd64-refined/tree/main), move to : model/clipseg

-<!-- ### Workflow
-[Workflow](./workflow.md) -->
+- [Download TripoSR](https://huggingface.co/stabilityai/TripoSR/blob/main/model.ckpt) and place it in `models/triposr`

+- [Download facebook/dino-vitb16](https://huggingface.co/facebook/dino-vitb16/tree/main) and place it in `models/triposr/facebook/dino-vitb16`

+[Download rembg Models](https://github.com/danielgatis/rembg/tree/main#Models),move to:`models/rembg`
+
+[Download lama](https://github.com/enesmsahin/simple-lama-inpainting/releases/download/v0.1.0/big-lama.pt), move to : `models/lama`
+
+[Download Salesforce/blip-image-captioning-base](https://huggingface.co/Salesforce/blip-image-captioning-base), move to :`models/clip_interrogator/Salesforce/blip-image-captioning-base`
+
+[Download succinctly/text2image-prompt-generator](https://huggingface.co/succinctly/text2image-prompt-generator/tree/main),move to:`models/prompt_generator/text2image-prompt-generator`
+
+[Download Helsinki-NLP/opus-mt-zh-en](https://huggingface.co/Helsinki-NLP/opus-mt-zh-en/tree/main),move to:`models/prompt_generator/opus-mt-zh-en`

 ## Installation

@@ -119,33 +349,51 @@ git clone https://github.com/shadowcz007/comfyui-mixlab-nodes.git
 Install the requirements:

 run directly:
+
 ```
 cd ComfyUI/custom_nodes/comfyui-mixlab-nodes
 install.bat
 ```

 or install the requirements using:
+
 ```
 ../../../python_embeded/python.exe -s -m pip install -r requirements.txt
 ```

 If you are using a venv, make sure you have it activated before installation and use:
+
 ```
 pip3 install -r requirements.txt
 ```

 #### Chinese community
-访问 [www.mixcomfy.com](https://www.mixcomfy.com)，获得更多内测功能，关注微信公众号：Mixlab无界社区

+访问 [www.mixcomfy.com](https://www.mixcomfy.com)，获得更多内测功能，关注微信公众号：Mixlab 无界社区

+####

-#### Thanks:
-[ComfyUI-CLIPSeg](https://github.com/biegert/ComfyUI-CLIPSeg/tree/main)
+File / LoadImagesFromPath SaveImageToLocal LoadImagesFromURL

 #### discussions:
+
 [discussions](https://github.com/shadowcz007/comfyui-mixlab-nodes/discussions)

-### TODO:
- 音频播放节点：带可视化、支持多音轨、可配置音轨音量
- vector https://github.com/GeorgLegato/stable-diffusion-webui-vectorstudio
-
+<picture>
+  <source
+    media="(prefers-color-scheme: dark)"
+    srcset="
+      https://api.star-history.com/svg?repos=shadowcz007/comfyui-mixlab-nodes&type=Date&theme=dark
+    "
+  />
+  <source
+    media="(prefers-color-scheme: light)"
+    srcset="
+      https://api.star-history.com/svg?repos=shadowcz007/comfyui-mixlab-nodes&type=Date
+    "
+  />
+  <img
+    alt="Star History Chart"
+    src="https://api.star-history.com/svg?repos=shadowcz007/comfyui-mixlab-nodes&type=Date"
+  />
+</picture>
@@ -0,0 +1,30 @@
+Jony Ive
+Dieter Rams
+Philippe Starck
+Karim Rashid
+Yves Béhar
+Marc Newson
+Naoto Fukasawa
+Jonathan Adler
+Patricia Urquiola
+Ross Lovegrove
+Tom Dixon
+Jasper Morrison
+Charles Eames
+Ray Eames
+Achille Castiglioni
+Ron Arad
+Konstantin Grcic
+Marcel Wanders
+Maarten Baas
+Stefan Sagmeister
+Ingo Maurer
+Hella Jongerius
+Sam Hecht
+Kim Colin
+Jaime Hayon
+Michael Anastassiades
+Nendo
+Oki Sato
+Matali Crasset
+Tokujin Yoshioka
@@ -0,0 +1,10 @@
+Chibi Anime Style
+Gakuen Anime Style
+Gekiga Anime Style
+Jidaimono Anime Style
+Kawaii Anime Style
+Mecha Anime Style
+Realistic Anime Style
+Semi-Realistic Anime Style
+Shoji Anime Style
+Kemonomimi Anime Style
@@ -0,0 +1,23 @@
+GoPro
+Drone
+polaroid
+black and white film
+Kodachrome
+shot on 8mm
+shot on 16mm
+shot on 35mm
+Microscopic
+Fisheye Lens
+Wide Angle
+Ultra-Wide Angle
+Panorama
+Short Exposure
+Long Exposure
+Double Exposure
+f2.8
+Depth of Field
+Soft Focus
+Deep Focus
+Shallow Focus
+Vanishing Point
+Vantage Point
@@ -0,0 +1,30 @@
+Elegant evening gown
+Casual jeans and t-shirt
+Formal black suit
+Stylish leather jacket
+Flowy bohemian dress
+Sporty tracksuit
+Chic little black dress
+Trendy ripped jeans
+Classic white button-down shirt
+Cozy oversized sweater
+Sophisticated tailored blazer
+Quirky patterned leggings
+Striped sailor top
+Polished knee-length skirt
+Vintage-inspired floral dress
+Edgy motorcycle jacket
+Preppy polo shirt
+Boho maxi skirt
+Professional pinstripe suit
+Relaxed denim shorts
+Glamorous sequined dress
+Athletic running shoes
+Formal bow tie
+Casual baseball cap
+Stylish fedora hat
+Warm woolen scarf
+Comfortable cotton socks
+Trendy ankle boots
+Cute summer sandals
+Cozy pajama set
@@ -0,0 +1,30 @@
+Happy
+Sad
+Angry
+Surprised
+Excited
+Worried
+Confused
+Disgusted
+Amused
+Bored
+Curious
+Embarrassed
+Frustrated
+Nervous
+Pleased
+Relieved
+Shy
+Tired
+Serious
+Silly
+Proud
+Grumpy
+Smug
+Sarcastic
+Flirty
+Skeptical
+Shocked
+Blissful
+Envious
+Mischievous
@@ -0,0 +1,10 @@
+[
+    {
+        "keyword":"Dog",
+        "imgurl":"http://127.0.0.1:8188/view?filename=1709966910233.png&type=input&subfolder=&rand=0.2734446552394221"
+    },
+    {
+        "keyword":"x",
+        "imgurl":"http://127.0.0.1:8188/view?filename=image%20(33).png&type=input&subfolder=pasted&rand=0.6984318219852814"
+    }
+]
@@ -0,0 +1,16 @@
+Mood Lighting
+Moody Lighting
+Studio Lighting
+Cove Lighting
+Soft Lighting
+Hard Lighting
+Volumetric Lighting
+Low-Key Lighting
+High-Key Lighting
+Epic Light
+Rembrandt Lighting
+Contre-Jour
+Veiling Flare
+Crepuscular Rays
+Rays of Shimmering Light
+Godrays
@@ -0,0 +1 @@
+{}
@@ -0,0 +1,132 @@
+Aaron Siskind
+Alessio Albi
+Alfred Eisenstaedt
+Alfred Stieglitz
+Alyssa Monks
+André Kertész
+Andreas Gursky
+Andrew Wyeth
+Anne Geddes
+Annie Leibovitz
+Ansel Adams
+Arnold Newman
+August Sander
+Balthus
+Berenice Abbott
+Bill Brandt
+Bill Henson
+Brassaï (Gyula Halász)
+Brooke Shaden
+Bruce Davidson
+Bruce Weber
+Bunny Yeager
+Carleton Watkins
+Carrie Mae Weems
+Chuck Close
+Cindy Sherman
+Clarence H. White
+Claude Cahun
+Danny Lyon
+David LaChapelle
+Dawoud Bey
+Diane Arbus
+Don McCullin
+Dora Maar
+Dorothea Lange
+Duane Michals
+Eadweard Muybridge
+Edward Burtynsky
+Edward Curtis
+Edward Ruscha
+Edward Steichen
+Edward Weston
+Elliott Erwitt
+Ernst Haas
+Eugene Atget
+Fan Ho
+Francesca Woodman
+Frans Lanting
+Garry Winogrand
+Georges Melies
+Gerda Taro
+Gertrude Käsebier
+Gordon Parks
+Graciela Iturbide
+Gregory Crewdson
+Harold Edgerton
+Helen Levitt
+Helmut Newton
+Hendrik Kerstens
+Henri Cartier-Bresson
+Hugh Kretschmer
+Irving Penn
+Jacques Henri Lartigue
+James Nachtwey
+James Van Der Zee
+Jay Maisel
+Jerry Uelsmann
+Joel Peter Witkin
+Joel Sartore
+John Frederick William Herschel
+Josef Sudek
+Julia Margaret Cameron
+Karl Blossfeldt
+Larry Burrows
+László Moholy-Nagy (photography)
+Lee Jeffries
+Lewis Hine
+Lorna Simpson
+Lynsey Addario
+Margaret Bourke-White
+Mario Testino
+Martin Parr
+Martin Schoeller
+Mary Ellen Mark
+Mathew B. Brady
+Méret Oppenheim
+Meryl McMaster
+Mick Rock
+Miles Aldridge
+Minor Martin White
+Nan Goldin
+Nathan Wirth
+Olive Cotton
+Olivier Rousteing
+Patrick Demarchelier
+Paul Nicklen
+Paul Outerbridge
+Paul Strand
+Pete Souza
+Peter Dombrovskis
+Peter Henry Emerson
+Peter Lik
+Peter Lindbergh
+Philip-Lorca diCorcia
+Philippe Halsman
+Ralph Gibson
+Richard Avedon
+Robert Adams
+Robert Bechtle
+Robert Capa
+Robert Frank
+Robert Mapplethorpe
+Roger Fenton
+Ruth Bernhard
+Sally Mann
+Sebastião Salgado
+Shirin Neshat
+Stefan Gesell
+Steven Meisel
+Susan Meiselas
+Vivian Maier
+Vivian Maier
+Viviane Sassen
+Walker Evans
+Wes Anderson
+William Eggleston
+William Eugene Smith
+William Henry Fox Talbot
+Yinka Shonibare
+Yousuf Karsh
+Man Ray
+Robert Mapplethorpe
@@ -0,0 +1,101 @@
+Doctor
+Teacher
+Engineer
+Lawyer
+Accountant
+Nurse
+Architect
+Chef
+Pilot
+Scientist
+Artist
+Writer
+Musician
+Actor
+Photographer
+Police officer
+Firefighter
+Dentist
+Pharmacist
+Veterinarian
+Electrician
+Plumber
+Carpenter
+Mechanic
+Farmer
+Astronaut
+Athlete
+Journalist
+Politician
+Economist
+Psychologist
+Social worker
+Librarian
+Translator
+Salesperson
+Entrepreneur
+Financial advisor
+Graphic designer
+Web developer
+Marketing manager
+Human resources manager
+Project manager
+Event planner
+Fashion designer
+Interior decorator
+Real estate agent
+Archaeologist
+Biologist
+Chemist
+Geologist
+Physicist
+Mathematician
+Historian
+Geographer
+Economist
+Sociologist
+Anthropologist
+Archaeologist
+Linguist
+Philosopher
+Economist
+Sociologist
+Anthropologist
+Archaeologist
+Linguist
+Philosopher
+Geographer
+Historian
+Economist
+Sociologist
+Anthropologist
+Archaeologist
+Linguist
+Philosopher
+Geographer
+Historian
+Economist
+Sociologist
+Anthropologist
+Archaeologist
+Linguist
+Philosopher
+Geographer
+Historian
+Economist
+Sociologist
+Anthropologist
+Archaeologist
+Linguist
+Philosopher
+Geographer
+Historian
+Economist
+Sociologist
+Anthropologist
+Archaeologist
+Linguist
+Philosopher
+Geographer
+Historian
+                     #MixCopilot
@@ -0,0 +1,58 @@
+Residential space
+Apartment building
+Villa
+Bungalow
+Condominium
+Commercial space
+Shopping mall
+Supermarket
+Restaurant
+Store
+Market
+Office space
+Office building
+Office
+Meeting room
+Co-working space
+Educational space
+School
+University
+Training institution
+Library
+Laboratory
+Medical space
+Hospital
+Clinic
+Pharmacy
+Nursing home
+Rehabilitation center
+Cultural space
+Museum
+Library
+Theater
+Concert hall
+Gallery
+Sports space
+Sports stadium
+Gym
+Swimming pool
+Basketball court
+Football field
+Transportation space
+Airport
+Train station
+Subway station
+Bus stop
+Parking lot
+Public space
+Park
+Square
+Street
+Pedestrian street
+Community center
+Industrial space
+Factory
+Warehouse
+Production workshop
+Mine
+Power plant
@@ -0,0 +1,135 @@
+Vintage
+Grain
+Sepia
+High Key
+Low Key
+High Dynamic Range
+Cross Process
+Radial Blur
+Infrared
+Lomo
+Photocopy
+Pencil Sketch
+Pop Art
+Orton
+Mosaic
+Selective Black and White
+Torn Paper
+Tilt-Shift
+Double Exposure
+Polaroid
+Liquid Ink
+Color Splash
+Sketch
+Water Drops
+Polarizer
+Chinese Painting
+Water Droplets
+Polarization
+Color Inversion
+Fish-eye
+Soft Focus
+Solarization
+Posterize
+Comic Book
+Duotone
+Gradient Map
+Edge Detection
+Oil Painting
+Reflection
+Mirror
+ASCII Art
+Glitch
+Time-Lapse
+Day to Night
+Surreal
+Black and White
+Sepia Tone
+Vintage Film
+Grainy Texture
+High Key Lighting
+Low Key Lighting
+Cross Processed Film
+Infrared Photography
+Photocopy
+Pencil Drawing
+Pop Art Filter
+Mosaic Filter
+Selective Desaturation
+Torn Paper
+Tilt-Shift Photography
+Double Exposure
+Polaroid Style Frame
+Water Drops Texture
+Polarizer
+Chinese Painting
+Water Droplets Texture
+Polarization
+Color Inversion
+Fish-eye Lens
+Soft Focus
+Solarize Filter
+Edge Detection
+Oil Painting
+Reflection
+Mirror Image
+Time-Lapse Photography
+Day to Night Transition
+Surreal Art Style
+Abstract Expressionism
+Acrylic Painting
+Anime
+Art Deco
+Biomorphic Abstraction
+Black and White Photograph
+Cartoon
+Charcoal Sketch
+Chibi Anime
+Chinese Painting
+Classicist Painting
+Collage
+Concept Art
+Cyberpunk
+Dada Art
+Digital Art
+Fantasy Art
+Fashion Art
+Fashion Sketch
+Fish-Eye lens Photograph
+Goth Art
+Graffiti
+Harlem Renaissance
+High Key Photograph
+Hyperrealist Pencil Sketch
+Impressionist Painting
+Josei Anime
+Long Exposure Photograph
+Low Key Photograph
+Macro Photograph
+Manga
+Metal Sculpture
+Mid Century Modern Illustration
+Mixed Media
+Modern Art
+Moe Anime
+Nihonga
+Origami
+Paper Mache
+Pen and Ink
+Pencil Sketch
+Photograph
+Photorealism
+Pinup Art
+Romanticist Painting
+Sci-Fi Art
+Semi Realistic Fantasy Art
+Semi Realistic Cyberpunk Art
+Shallow Depth of Field Photograph
+Steam Punk Art
+Stone Sculpture
+Superhero Comic
+Surrealist Art
+Tempura Painting
+Underground Comic
+Watercolor Painting
+Zulu Urban Art
@@ -10,6 +10,12 @@ if exist "%python_exec%" (
    for /f "delims=" %%i in (%requirements_txt%) do (
        %python_exec% -s -m pip install "%%i" -i https://pypi.tuna.tsinghua.edu.cn/simple
    )
+
+    @REM %python_exec% -s -m pip install --upgrade --force llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu121
+
+    @REM %python_exec% -s -m pip install --upgrade --force llama-cpp-python[server]
+    
+
 ) else (
    echo Installing with system Python
    for /f "delims=" %%i in (%requirements_txt%) do (
@@ -1,12 +1,23 @@

-
-
+import os
+import folder_paths
+import torchaudio

 class SpeechRecognition:
    @classmethod
    def INPUT_TYPES(s):
        return {"required": { 
                    "upload":("AUDIOINPUTMIX",),   },
+                "optional":{
+                    "start_by":("INT", {
+                        "default": 0, 
+                        "min": 0, #Minimum value
+                        "max": 2048, #Maximum value
+                        "step": 1, #Slider's step
+                        "display": "number" # Cosmetic only: display as "number" or "slider"
+                    }),
+                   
+                }
                }
    
    RETURN_TYPES = ("STRING",)
@@ -14,13 +25,13 @@ class SpeechRecognition:

    FUNCTION = "run"

-    CATEGORY = "♾️Mixlab/audio"
+    CATEGORY = "♾️Mixlab/Audio"

    INPUT_IS_LIST = False
    OUTPUT_IS_LIST = (False,)

-    def run(self,upload):
-        return (upload,)
+    def run(self,upload,start_by):
+        return {"ui": {"start_by": [start_by]}, "result": (upload,)}


 class SpeechSynthesis:
@@ -38,9 +49,72 @@ class SpeechSynthesis:
    OUTPUT_NODE = True
    OUTPUT_IS_LIST = (True,)

-    CATEGORY = "♾️Mixlab/audio"
+    CATEGORY = "♾️Mixlab/Audio"

    def run(self, text):
        # print(session_history)
        return {"ui": {"text": text}, "result": (text,)}
    
+
+
+class AudioPlayNode:
+    def __init__(self):
+        self.output_dir = folder_paths.get_temp_directory()
+        self.type = "temp"
+        self.prefix_append = ""
+        self.compress_level = 4
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "audio": ("AUDIO",),
+              }, 
+                }
+    
+    RETURN_TYPES = ()
+  
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Audio"
+
+    INPUT_IS_LIST = False
+    OUTPUT_IS_LIST = ()
+
+    OUTPUT_NODE = True
+  
+    def run(self,audio):
+
+        # 判断是否是 Tensor 类型
+        is_tensor = not isinstance(audio, dict)
+        # print('#判断是否是 Tensor 类型',is_tensor,audio)
+        if not is_tensor and 'waveform' in audio and 'sample_rate' in audio:
+            # {'waveform': tensor([], size=(1, 1, 0)), 'sample_rate': 44100}
+            is_tensor=True
+
+        if is_tensor and (not 'audio_path' in audio):
+            filename_prefix=""
+            # 保存
+            filename_prefix += self.prefix_append
+            full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, self.output_dir)
+            results = list()
+            
+            filename_with_batch_num = filename.replace("%batch_num%", str(1))
+            file = f"{filename_with_batch_num}_{counter:05}_.wav"
+            
+            torchaudio.save(os.path.join(full_output_folder, file), audio['waveform'].squeeze(0), audio["sample_rate"])
+            results.append({
+                    "filename": file,
+                    "subfolder": subfolder,
+                    "type": self.type
+                })
+            
+        else:
+            results=[{
+                    "filename": audio['filename'],
+                    "subfolder":audio['subfolder'],
+                    "type": audio['type'],
+                    "audio_path":audio['audio_path']
+                }]
+                
+
+        # print(audio)
+        return {"ui": {"audio":results}}
@@ -1,7 +1,92 @@
 import openai 
 import time
 import urllib.error
-import re,json
+import re,json,os,string,random
+import folder_paths
+import hashlib
+import codecs,sys
+import importlib.util
+import subprocess
+
+python = sys.executable
+
+# 从文本中提取json
+def extract_json_strings(text):
+    json_strings = []
+    brace_level = 0
+    json_str = ''
+    in_json = False
+    
+    for char in text:
+        if char == '{':
+            brace_level += 1
+            in_json = True
+        if in_json:
+            json_str += char
+        if char == '}':
+            brace_level -= 1
+        if in_json and brace_level == 0:
+            json_strings.append(json_str)
+            json_str = ''
+            in_json = False
+
+    return json_strings[0] if len(json_strings)>0 else "{}"
+
+
+def is_installed(package, package_overwrite=None,auto_install=True):
+    is_has=False
+    try:
+        spec = importlib.util.find_spec(package)
+        is_has=spec is not None
+    except ModuleNotFoundError:
+        pass
+
+    package = package_overwrite or package
+
+    if spec is None:
+        if auto_install==True:
+            print(f"Installing {package}...")
+            # 清华源 -i https://pypi.tuna.tsinghua.edu.cn/simple
+            command = f'"{python}" -m pip install {package}'
+    
+            result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, shell=True, env=os.environ)
+
+            is_has=True
+
+            if result.returncode != 0:
+                print(f"Couldn't install\nCommand: {command}\nError code: {result.returncode}")
+                is_has=False
+    else:
+        print(package+'## OK')
+
+    return is_has
+  
+
+
+# def is_installed(package):
+#     try:
+#         spec = importlib.util.find_spec(package)
+#     except ModuleNotFoundError:
+#         return False
+#     return spec is not None
+
+
+def get_unique_hash(string):
+    hash_object = hashlib.sha1(string.encode())
+    unique_hash = hash_object.hexdigest()
+    return unique_hash
+
+def generate_random_string(length):
+    letters = string.ascii_letters + string.digits
+    return ''.join(random.choice(letters) for _ in range(length))
+
+class AnyType(str):
+  """A special class that is always equal in not equal comparisons. Credit to pythongosssss"""
+
+  def __ne__(self, __value: object) -> bool:
+    return False
+
+any_type = AnyType("*")

 #  判断是否是azure服务
 def is_azure_url(url):
@@ -23,23 +108,123 @@ def azure_client(key,url):

 def openai_client(key,url):
    client = openai.OpenAI(
-    api_key=key,
-    base_url=url
+        api_key=key,
+        base_url=url
    )
    return client

+def ZhipuAI_client(key):
+    try:
+        if is_installed('zhipuai')==True:
+            from zhipuai import ZhipuAI
+    except:
+        print("#install zhipuai error")
+
+    client = ZhipuAI(
+        api_key=key, # 填写您的 APIKey
+    ) 
+    return client
+
+
+# 优先使用phi
+def phi_sort(lst):
+    return sorted(lst, key=lambda x: x.lower().count('phi'), reverse=True)
+
+def get_llama_path():
+    try:
+        return folder_paths.get_folder_paths('llamafile')[0]
+    except:
+        return os.path.join(folder_paths.models_dir, "llamafile")
+
+# def get_llama_models():
+#     res=[]
+
+#     model_path=get_llama_path()
+#     if os.path.exists(model_path):
+#         files = os.listdir(model_path)
+#         for file in files:
+#             if os.path.isfile(os.path.join(model_path, file)):
+#                 res.append(file)
+#         res=phi_sort(res)
+#     return res
+
+# llama_modes_list=get_llama_models()
+# llama_modes_list=[]
+
+# def get_llama_model_path(file_name):
+#     model_path=get_llama_path()
+#     mp=os.path.join(model_path,file_name)
+#     return mp
+
+# def llama_cpp_client(file_name):
+#     try:
+#         if is_installed('llama_cpp')==False:
+#             import subprocess
+
+#             # 安装
+#             print('#pip install llama-cpp-python')
+ 
+#             result = subprocess.run([sys.executable, '-s', '-m', 'pip', 
+#                                      'install', 
+#                                      'llama-cpp-python',
+#                                      '--extra-index-url',
+#                                      'https://abetlen.github.io/llama-cpp-python/whl/cu121'
+#                                      ], capture_output=True, text=True)
+
+#             #检查命令执行结果
+#             if result.returncode == 0:
+#                 print("#install success")
+#                 from llama_cpp import Llama
+
+#                 subprocess.run([sys.executable, '-s', '-m', 'pip', 
+#                                      'install', 
+#                                      'llama-cpp-python[server]'
+#                                      ], capture_output=True, text=True)
+
+#             else:
+#                 print("#install error")
+            
+#         else:
+#             from llama_cpp import Llama
+#     except:
+#         print("#install llama-cpp-python error")
+
+#     if file_name:
+#         mp=get_llama_model_path(file_name)
+#         # file_name=get_llama_models()[0]
+#         # model_path=os.path.join(folder_paths.models_dir, "llamafile")
+#         # mp=os.path.join(model_path,file_name)
+
+#         llm = Llama(model_path=mp, chat_format="chatml",n_gpu_layers=-1,n_ctx=512)
+
+#         return llm
+
+
+if is_installed('json_repair'):
+    from json_repair import repair_json


 def chat(client, model_name,messages ):
-
+        print('#chat',model_name,messages)
        try_count = 0
        while True:
            try_count += 1
            try:
-                response = client.chat.completions.create(
-                    model=model_name,
-                    messages=messages
-                )
+                if hasattr(client, "chat"):
+                    response = client.chat.completions.create(
+                        model=model_name,
+                        messages=messages
+                    )
+                else:
+                    # 是llama的
+                    response = client.create_chat_completion_openai_v1(
+                        messages=messages,
+                        # response_format={
+                        #     "type": "json_object",
+                        # },
+                        # temperature=0.7,
+                    )
+
                break
            except openai.AuthenticationError as ex:
                raise ex
@@ -48,7 +233,8 @@ def chat(client, model_name,messages ):
                    raise ex
                time.sleep(3)
                continue
-
+        
+        # print(response.keys())
        finish_reason = response.choices[0].finish_reason
        if finish_reason != "stop":
            raise RuntimeError("API finished with unexpected reason: " + finish_reason)
@@ -62,6 +248,36 @@ def chat(client, model_name,messages ):
        return content


+llm_apis=[
+            {
+                "value": "https://api.openai.com/v1",
+                "label": "openai"
+            },
+            {
+                "value": "https://openai.api2d.net/v1",
+                "label": "api2d"
+            },
+            # {
+            #     "value": "https://docs-test-001.openai.azure.com",
+            #     "label": "https://docs-test-001.openai.azure.com"
+            # },
+             
+            {
+                "value": "https://api.moonshot.cn/v1",
+                "label": "Kimi"
+            },
+            {
+                "value": "https://api.deepseek.com/v1",
+                "label": "DeepSeek-V2"
+            },
+            {
+                "value": "https://api.siliconflow.cn/v1",
+                "label": "SiliconCloud"
+            }]
+        
+llm_apis_dict = {api["label"]: api["value"] for api in llm_apis}
+
+
 class ChatGPTNode:
    def __init__(self):
        # self.__client = OpenAI()
@@ -71,25 +287,60 @@ class ChatGPTNode:

    @classmethod
    def INPUT_TYPES(cls):
+
+        model_list=[ 
+            "gpt-3.5-turbo",
+            "gpt-3.5-turbo-16k",
+            "gpt-4o",
+            "gpt-4o-2024-05-13",
+            "gpt-4",
+            "gpt-4-0314",
+            "gpt-4-0613",
+            "gpt-3.5-turbo-0301",
+            "gpt-3.5-turbo-0613",
+            "gpt-3.5-turbo-16k-0613",
+            "qwen-turbo",
+            "qwen-plus",
+            "qwen-long",
+            "qwen-max",
+            "qwen-max-longcontext",
+            "glm-4",
+            "glm-3-turbo",
+            "moonshot-v1-8k",
+            "moonshot-v1-32k",
+            "moonshot-v1-128k",
+            "deepseek-chat",
+            "Qwen/Qwen2-7B-Instruct",
+            "THUDM/glm-4-9b-chat",
+            "01-ai/Yi-1.5-9B-Chat-16K",
+            "meta-llama/Meta-Llama-3.1-8B-Instruct"
+                    ]
+        
        return {
            "required": {
-                "api_key":("KEY", {"default": "", "multiline": True}),
-                "api_url":("URL", {"default": "", "multiline": True}),
-                "prompt": ("STRING", {"multiline": True}),
+                # "api_key":("KEY", {"default": "", "multiline": True,"dynamicPrompts": False}),
+                # "api_key":("STRING", {"forceInput": True,}),
+               
+                "prompt": ("STRING", {"multiline": True,"dynamicPrompts": False}),
                "system_content": ("STRING", 
                                   {
                                       "default": "You are ChatGPT, a large language model trained by OpenAI. Answer as concisely as possible.", 
-                                       "multiline": True
+                                       "multiline": True,"dynamicPrompts": False
                                       }),
-                "model": (["gpt-3.5-turbo","gpt-35-turbo","gpt-3.5-turbo-16k", "gpt-3.5-turbo-16k-0613", "gpt-4-0613","gpt-4-1106-preview"], 
-                          {"default": "gpt-3.5-turbo"}),
-                "seed": ("INT", {"default": 0, "min": 0, "max": 10000, "step": 1}),
+                 
+                "model": ( model_list, 
+                    {"default": model_list[0]}),
+                "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff, "step": 1}),
                "context_size":("INT", {"default": 1, "min": 0, "max":30, "step": 1}),
+                "api_url":(list(llm_apis_dict.keys()), 
+                    {"default": list(llm_apis_dict.keys())[0]}),
            },
-             "hidden": {
-                "unique_id": "UNIQUE_ID",
-                "extra_pnginfo": "EXTRA_PNGINFO",
-            },
+             "optional":{
+                    "api_key":("STRING", {"forceInput": True,}),
+                    "custom_model_name":("STRING", {"forceInput": True,}), #适合自定义model
+                     "custom_api_url":("STRING", {"forceInput": True,}), #适合自定义model
+                },
+             
        }

    RETURN_TYPES = ("STRING","STRING","STRING",)
@@ -101,12 +352,29 @@ class ChatGPTNode:

    
    def generate_contextual_text(self,
-                                 api_key,
-                                 api_url, 
+                                #  api_key,
                                 prompt, 
                                 system_content,
-                                   model, 
-                                   seed,context_size,unique_id = None, extra_pnginfo=None):
+                                 model, 
+                                seed,
+                                context_size,
+                                api_url,
+                                api_key=None,
+                                custom_model_name=None,
+                                custom_api_url=None,
+                                ):
+        
+        if custom_model_name!=None:
+            model=custom_model_name
+
+        api_url=llm_apis_dict[api_url] if api_url in llm_apis_dict else ""
+
+        if custom_api_url!=None:
+            api_url=custom_api_url
+
+        if api_key==None:
+            api_key="lm_studio"
+
        # print(api_key!='',api_url,prompt,system_content,model,seed)
        # 可以选择保留会话历史以维持上下文记忆
        # 或者在此处清除会话历史 self.session_history.clear()
@@ -119,13 +387,21 @@ class ChatGPTNode:
            self.system_content=system_content
            # self.session_history=[]
            # self.session_history.append({"role": "system", "content": system_content})
-        
+        print("api_key,api_url",api_key,api_url)
        # 
        if is_azure_url(api_url):
            client=azure_client(api_key,api_url)
        else:
-            client=openai_client(api_key,api_url)
-            print('openai url')
+            # 根据用户选择的模型，设置相应的接口和模型名称
+            if model == "glm-4" :
+                client = ZhipuAI_client(api_key)  # 使用 Zhipuai 的接口
+                print('using Zhipuai interface')
+            # elif model in llama_modes_list:
+            #     #
+            #     client=llama_cpp_client(model)
+            else :
+                client = openai_client(api_key,api_url)  # 使用 ChatGPT  的接口
+                # print('using ChatGPT interface',api_key,api_url)

        # 把用户的提示添加到会话历史中
        # 调用API时传递整个会话历史
@@ -141,6 +417,7 @@ class ChatGPTNode:
        session_history=crop_list_tail(self.session_history,context_size)

        messages=[{"role": "system", "content": self.system_content}]+session_history+[{"role": "user", "content": prompt}]
+
        response_content = chat(client,model,messages)
        
        self.session_history=self.session_history+[{"role": "user", "content": prompt}]+[{'role':'assistant',"content":response_content}]
@@ -161,14 +438,104 @@ class ChatGPTNode:
        return (response_content,json.dumps(messages, indent=4),json.dumps(self.session_history, indent=4),)


+class SiliconflowFreeNode:
+    def __init__(self):
+        # self.__client = OpenAI()
+        self.session_history = []  # 用于存储会话历史的列表
+        # self.seed=0
+        self.system_content="You are ChatGPT, a large language model trained by OpenAI. Answer as concisely as possible."
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        model_list= [ 
+            "Qwen/Qwen2-7B-Instruct",
+            "THUDM/glm-4-9b-chat",
+            "01-ai/Yi-1.5-9B-Chat-16K",
+            "meta-llama/Meta-Llama-3.1-8B-Instruct"
+            ]
+        return {
+            "required": {
+                "api_key":("STRING", {"forceInput": True,}),
+                "prompt": ("STRING", {"multiline": True,"dynamicPrompts": False}),
+                "system_content": ("STRING", 
+                                   {
+                                       "default": "You are ChatGPT, a large language model trained by OpenAI. Answer as concisely as possible.", 
+                                       "multiline": True,"dynamicPrompts": False
+                                       }),
+                "model": ( model_list, 
+                    {"default": model_list[0]}),
+                "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff, "step": 1}),
+                "context_size":("INT", {"default": 1, "min": 0, "max":30, "step": 1}),
+            },
+               "optional":{
+                    "custom_model_name":("STRING", {"forceInput": True,}), #适合自定义model
+                }, 
+        }
+
+    RETURN_TYPES = ("STRING","STRING","STRING",)
+    RETURN_NAMES = ("text","messages","session_history",)
+    FUNCTION = "generate_contextual_text"
+    CATEGORY = "♾️Mixlab/GPT"
+    INPUT_IS_LIST = False
+    OUTPUT_IS_LIST = (False,False,False,)
+
+    
+    def generate_contextual_text(self,
+                                 api_key,
+                                 prompt, 
+                                 system_content,
+                                model, 
+                                seed,context_size,custom_model_name=None):
+
+        if custom_model_name!=None:
+            model=custom_model_name
+
+        api_url="https://api.siliconflow.cn/v1"
+        
+        # 把系统信息和初始信息添加到会话历史中
+        if system_content:
+            self.system_content=system_content
+            # self.session_history=[]
+            # self.session_history.append({"role": "system", "content": system_content})
+        
+        # 
+        client = openai_client(api_key,api_url)  # 使用 ChatGPT  的接口
+        # print('using ChatGPT interface',api_key,api_url)
+
+        # 把用户的提示添加到会话历史中
+        # 调用API时传递整个会话历史
+
+        def crop_list_tail(lst, size):
+            if size >= len(lst):
+                return lst
+            elif size==0:
+                return []
+            else:
+                return lst[-size:]
+            
+        session_history=crop_list_tail(self.session_history,context_size)
+
+        messages=[{"role": "system", "content": self.system_content}]+session_history+[{"role": "user", "content": prompt}]
+
+        response_content = chat(client,model,messages)
+        
+        self.session_history=self.session_history+[{"role": "user", "content": prompt}]+[{'role':'assistant',"content":response_content}]
+
+        return (response_content,json.dumps(messages, indent=4),json.dumps(self.session_history, indent=4),)
+
+
+

 class ShowTextForGPT:
    @classmethod
    def INPUT_TYPES(s):
        return {
            "required": {
-                "text": ("STRING", {"forceInput": True}),
-            }
+                "text": ("STRING", {"forceInput": True,"dynamicPrompts": False}),
+            },
+            "optional":{ 
+                    "output_dir": ("STRING",{"forceInput": True,"default": "","multiline": True,"dynamicPrompts": False}), 
+                }
        }

    INPUT_IS_LIST = True
@@ -177,11 +544,65 @@ class ShowTextForGPT:
    OUTPUT_NODE = True
    OUTPUT_IS_LIST = (True,)

-    CATEGORY = "♾️Mixlab/GPT"
+    CATEGORY = "♾️Mixlab/Text"

-    def run(self, text):
-        # print(session_history)
+    def run(self, text,output_dir=[""]):
+        
+        # 类型纠正
+        texts=[]
+        for t in text:
+            if not isinstance(t, str):
+                t = str(t)
+            texts.append(t)
+
+        text=texts
+
+        if len(output_dir)==1 and (output_dir[0]=='' or os.path.dirname(output_dir[0])==''):
+            t='\n'.join(text)
+            output_dir=[
+                os.path.join(folder_paths.get_temp_directory(),
+                             get_unique_hash(t)+'.txt'
+                             )
+            ]
+        elif len(output_dir)==1:
+            base=os.path.basename(output_dir[0])
+            t='\n'.join(text)
+            if base=='' or os.path.splitext(base)[1]=='':
+                base=get_unique_hash(t)+'.txt'
+            output_dir=[
+                os.path.join(output_dir[0],
+                             base
+                             )
+            ]
+        # elif len(output_dir)>1:
+
+        
+
+        if len(output_dir)==1 and len(text)>1:
+            output_dir=[output_dir[0] for _ in range(len(text))]
+        
+        for i in range(len(text)):
+
+            o_fp=output_dir[i]
+            dirp=os.path.dirname(o_fp)
+            if dirp=='':
+                dirp=folder_paths.get_temp_directory()
+                o_fp=os.path.join(folder_paths.get_temp_directory(),o_fp
+                             )
+
+            if not os.path.exists(dirp):
+                os.mkdir(dirp)
+
+            if not os.path.splitext(o_fp)[1].lower()=='.txt':
+                o_fp=o_fp+'.txt'
+
+            t=text[i]
+            with open(o_fp, 'w') as file:
+                file.write(t)
+
+        # print(text)
        return {"ui": {"text": text}, "result": (text,)}
+        


 class CharacterInText:
@@ -189,8 +610,8 @@ class CharacterInText:
    def INPUT_TYPES(s):
        return {
            "required": {
-                 "text": ("STRING", {"multiline": True}),
-                 "character": ("STRING", {"multiline": True}),
+                 "text": ("STRING", {"multiline": True,"dynamicPrompts": False}),
+                 "character": ("STRING", {"multiline": True,"dynamicPrompts": False}),
                 "start_index": ("INT", {
                    "default": 1,
                    "min": 0, #Minimum value
@@ -207,11 +628,99 @@ class CharacterInText:
    # OUTPUT_NODE = True
    OUTPUT_IS_LIST = (False,)

-    CATEGORY = "♾️Mixlab/GPT"
+    CATEGORY = "♾️Mixlab/Text"

    def run(self, text,character,start_index):
        # print(text,character,start_index)
-        b=1 if character in text else 0
+        b=1 if character.lower() in text.lower() else 0
        
        return (b+start_index,)

+class TextSplitByDelimiter:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "text": ("STRING", {"multiline": True,"dynamicPrompts": False}),
+                "delimiter":("STRING", {"multiline": False,"default":",","dynamicPrompts": False}),
+                "start_index": ("INT", {
+                    "default": 0,
+                    "min": 0, #Minimum value
+                    "max": 1000, #Maximum value
+                    "step": 1, #Slider's step
+                    "display": "number" # Cosmetic only: display as "number" or "slider"
+                }),
+                 "skip_every": ("INT", {
+                    "default": 0,
+                    "min": 0, #Minimum value
+                    "max": 10, #Maximum value
+                    "step": 1, #Slider's step
+                    "display": "number" # Cosmetic only: display as "number" or "slider"
+                }),
+                "max_count": ("INT", {
+                    "default": 10,
+                    "min": 1, #Minimum value
+                    "max": 1000, #Maximum value
+                    "step": 1, #Slider's step
+                    "display": "number" # Cosmetic only: display as "number" or "slider"
+                }),
+            }
+        }
+
+    INPUT_IS_LIST = False
+    RETURN_TYPES = ("STRING",)
+    FUNCTION = "run"
+    # OUTPUT_NODE = True
+    OUTPUT_IS_LIST = (True,)
+
+    CATEGORY = "♾️Mixlab/Text"
+
+    def run(self, text,delimiter,start_index,skip_every,max_count):
+         
+        if delimiter=="":
+            arr=[text.strip()]
+        else:
+            delimiter=codecs.decode(delimiter, 'unicode_escape')
+            arr= [line for line in text.split(delimiter) if line.strip()]
+
+        arr= arr[start_index:start_index + max_count * (skip_every+1):(skip_every+1)]
+
+        return (arr,)
+
+
+class JsonRepair:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                 "json_string":("STRING", {"forceInput": True,}), 
+                 "key":("STRING", {"multiline": False,"dynamicPrompts": False,"default": ""}),
+            }
+        }
+
+    INPUT_IS_LIST = False
+    RETURN_TYPES = ("STRING","STRING",)
+    RETURN_NAMES = ("json_string","value",)
+    FUNCTION = "run"
+    # OUTPUT_NODE = True
+    OUTPUT_IS_LIST = (False,False,)
+
+    CATEGORY = "♾️Mixlab/GPT"
+
+    def run(self, json_string,key=""):
+
+        json_string=extract_json_strings(json_string)
+        # print(json_string)
+        good_json_string = repair_json(json_string)
+
+        # 将 JSON 字符串解析为 Python 对象
+        data = json.loads(good_json_string)
+
+        v=""
+        if key!="" and (key in data):
+            v=data[key]
+
+        # 将 Python 对象转换回 JSON 字符串，确保中文字符不被转义
+        json_str_with_chinese = json.dumps(data, ensure_ascii=False)
+
+        return (json_str_with_chinese,v,)
@@ -0,0 +1,283 @@
+import os,sys
+import folder_paths
+
+from PIL import Image
+import importlib.util
+
+import comfy.utils
+import numpy as np
+import json
+import torch
+import random
+
+
+# from clip_interrogator import Config, Interrogator
+
+
+global _available
+_available=False
+
+def is_installed(package):
+    try:
+        spec = importlib.util.find_spec(package)
+    except ModuleNotFoundError:
+        return False
+    return spec is not None
+
+
+try:
+    if is_installed('clip_interrogator')==False:
+        import subprocess
+
+        # 安装
+        print('#pip install clip-interrogator==0.6.0')
+
+        result = subprocess.run([sys.executable, '-s', '-m', 'pip', 'install', 'clip-interrogator==0.6.0'], capture_output=True, text=True)
+
+        #检查命令执行结果
+        if result.returncode == 0:
+            print("#install success")
+            from clip_interrogator import Config, Interrogator
+            _available=True
+        else:
+            print("#install error")
+        
+    else:
+        from clip_interrogator import Config, Interrogator
+        _available=True
+
+except:
+    _available=False
+
+try:
+    from transformers import AutoProcessor, BlipForConditionalGeneration
+except:
+    _available=False
+    print('pls check  transformers.__version__>=4.36.0:: AutoProcessor, BlipForConditionalGeneration')
+
+
+
+def load_caption_model(model_path,config,t='blip-base'):
+    dtype=torch.float16 if config.device == 'cuda' else torch.float32
+    caption_model = BlipForConditionalGeneration.from_pretrained(model_path, torch_dtype=dtype)
+    
+    caption_processor = AutoProcessor.from_pretrained(model_path)
+
+    caption_model.eval()
+    if not config.caption_offload:
+        caption_model = caption_model.to(config.device)
+    
+    return (caption_model,caption_processor)
+    
+
+def get_clip_interrogator_path():
+    try:
+        return folder_paths.get_folder_paths('clip_interrogator')[0]
+    except:
+        return os.path.join(folder_paths.models_dir, "clip_interrogator")
+
+
+cache_path=get_clip_interrogator_path()
+
+caption_model_path=os.path.join(cache_path, "Salesforce","blip-image-captioning-base")
+if not os.path.exists(caption_model_path):
+    print(f"## clip_interrogator_model not found: {caption_model_path}, pls download from https://huggingface.co/Salesforce/blip-image-captioning-base")
+    caption_model_path='Salesforce/blip-image-captioning-base'
+
+
+# Tensor to PIL
+def tensor2pil(image):
+    return Image.fromarray(np.clip(255. * image.cpu().numpy().squeeze(), 0, 255).astype(np.uint8))
+
+# Convert PIL to Tensor
+def pil2tensor(image):
+    return torch.from_numpy(np.array(image).astype(np.float32) / 255.0).unsqueeze(0)
+
+
+def image_analysis_fn(ci,image):
+    image = image.convert('RGB')
+    image_features = ci.image_to_features(image)
+
+    top_mediums = ci.mediums.rank(image_features, 5)
+    top_artists = ci.artists.rank(image_features, 5)
+    top_movements = ci.movements.rank(image_features, 5)
+    top_trendings = ci.trendings.rank(image_features, 5)
+    top_flavors = ci.flavors.rank(image_features, 5)
+
+    medium_ranks = {medium: sim for medium, sim in zip(top_mediums, ci.similarities(image_features, top_mediums))}
+    artist_ranks = {artist: sim for artist, sim in zip(top_artists, ci.similarities(image_features, top_artists))}
+    movement_ranks = {movement: sim for movement, sim in zip(top_movements, ci.similarities(image_features, top_movements))}
+    trending_ranks = {trending: sim for trending, sim in zip(top_trendings, ci.similarities(image_features, top_trendings))}
+    flavor_ranks = {flavor: sim for flavor, sim in zip(top_flavors, ci.similarities(image_features, top_flavors))}
+    
+    return {
+        "medium_ranks":medium_ranks, 
+        "artist_ranks":artist_ranks, 
+        "movement_ranks":movement_ranks, 
+        "trending_ranks":trending_ranks, 
+        "flavor_ranks":flavor_ranks
+        }
+
+
+def generate_sentences(data):
+    sentences = []
+
+    # Get the length of data
+    data_length = len(data)
+
+    # Use a recursive function to handle variable-length data
+    def generate_recursive(index, current_sentence, current_score):
+        # Check if recursion is complete
+        if index == data_length:
+            sentences.append({"sentence": current_sentence, "score": current_score})
+            return
+
+        # Get the current level data
+        current_data = data[index]
+
+        # Iterate through the current level data
+        for phrase in current_data:
+            sentence = current_sentence + ("," if current_sentence.strip() else "") + phrase
+            score = current_score + current_data[phrase]
+            generate_recursive(index + 1, sentence, score)
+
+    # Start recursive generation of sentences
+    generate_recursive(0, "", 0)
+
+    # Sort the generated sentences by score in descending order
+    sentences.sort(key=lambda x: x["score"], reverse=True)
+
+    def get_random_elements(elements, num):
+        return random.sample(elements, num)
+
+    ps = get_random_elements(sentences, 5)
+    ps = [s["sentence"] for s in sorted(ps, key=lambda x: x["score"], reverse=True)]
+
+    return ps
+
+
+
+
+def image_to_prompt(ci,image, mode):
+    ci.config.chunk_size = 2048 if ci.config.clip_model_name == "ViT-L-14/openai" else 1024
+    ci.config.flavor_intermediate_count = 2048 if ci.config.clip_model_name == "ViT-L-14/openai" else 1024
+    image = image.convert('RGB')
+    if mode == 'best':
+        return ci.interrogate(image)
+    elif mode == 'classic':
+        return ci.interrogate_classic(image)
+    elif mode == 'fast':
+        return ci.interrogate_fast(image)
+    elif mode == 'negative':
+        return ci.interrogate_negative(image)
+
+# image = Image.open(image_path).convert('RGB')
+# ci = Interrogator(Config(clip_model_name="ViT-L-14/openai"))
+# print(ci.interrogate(image))
+
+
+class ClipInterrogator:
+
+    global _available
+    available=_available
+    
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "image": ("IMAGE",),
+            "prompt_mode": (['fast','classic','best','negative'],),
+            "image_analysis": (["off","on"],), 
+                             },
+
+                # "optional":{
+                #     "output":("CLIPINTERROGATOR", {"multiline": True,"default": "", "dynamicPrompts": False})
+                # },
+
+                }
+    
+    RETURN_TYPES = ("STRING","STRING",)
+    RETURN_NAMES = ("prompt","random_samples",)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Prompt"
+    OUTPUT_NODE = True
+    INPUT_IS_LIST = True
+    OUTPUT_IS_LIST = (True,True,)
+    global ci
+    ci = None
+    def run(self,image,prompt_mode,image_analysis):
+        global ci
+
+        prompt_mode=prompt_mode[0]
+        analysis=image_analysis[0]
+
+        prompt_result=[]
+        analysis_result=[]
+
+        # 进度条
+        pbar = comfy.utils.ProgressBar(len(image)*(2 if analysis=='on' else 1))
+        
+        if ci==None:
+            config=Config(
+                clip_model_name="ViT-L-14/openai",
+                device="cuda" if torch.cuda.is_available() else "cpu",
+                download_cache=True,
+                clip_model_path=cache_path,
+                cache_path=cache_path
+                )
+            config.apply_low_vram_defaults()
+
+            caption_model,caption_processor=load_caption_model(caption_model_path,config)
+
+            config.caption_model= caption_model
+            config.caption_processor= caption_processor
+
+            ci = Interrogator(config)
+        # else:
+        #     simple_lama.model.to("cuda" if torch.cuda.is_available() else "cpu")
+
+        for i in range(len(image)):
+            im=image[i]
+
+            im=tensor2pil(im)
+            im=im.convert('RGB')
+
+            if analysis=='on':
+                analysis_res=image_analysis_fn(ci,im)
+                analysis_result.append( analysis_res )
+                pbar.update(1)
+
+            prompt=image_to_prompt(ci,im,prompt_mode)
+            pbar.update(1)
+            prompt_result.append(prompt)
+
+
+        # result.save("inpainted.png")
+        if ci.config.clip_offload and not ci.clip_offloaded:
+            ci.clip_model = ci.clip_model.to('cpu')
+            ci.clip_offloaded = True
+
+        if ci.config.caption_offload and not ci.caption_offloaded:
+            ci.caption_model = ci.caption_model.to('cpu')
+            ci.caption_offloaded = True
+
+        # analysis_result=[]
+        # items = app.graph.getNodeById(31).widgets[2].value["items"]
+        
+        random_samples=[]
+
+        for r in analysis_result:
+            random_sample = generate_sentences([r['medium_ranks'], r['artist_ranks'],r['movement_ranks'],r['trending_ranks'],r['flavor_ranks']])
+            for s in random_sample:
+                random_samples.append(s)
+        # print(len(random_samples))
+        # print('-----')
+        # print( random_samples)
+        return {
+            "ui":{
+                    "prompt": prompt_result,
+                    "analysis":analysis_result,
+                    "random_samples":random_samples
+                },
+            "result": (prompt_result,random_samples,)}
@@ -1,261 +0,0 @@
-#### Thanks:
-# [ComfyUI-CLIPSeg](https://github.com/biegert/ComfyUI-CLIPSeg/tree/main)
-
-from transformers import CLIPSegProcessor, CLIPSegForImageSegmentation
-
-from PIL import Image
-import torch
-import torchvision.transforms as T
-import numpy as np
-
-from torchvision.transforms.functional import to_pil_image
-import matplotlib.pyplot as plt
-import matplotlib.cm as cm
-
-
-import cv2
-
-from scipy.ndimage import gaussian_filter
-
-from typing import Optional, Tuple
-
-import warnings,os
-warnings.filterwarnings("ignore", category=UserWarning, module="torch")
-warnings.filterwarnings("ignore", category=UserWarning, module="safetensors")
-
-import folder_paths
-import logging
-
-logger = logging.getLogger('CLIPSeg nodes')
-
-clipseg_model_dir = os.path.join(folder_paths.models_dir, "clipseg")
-
-if not os.path.exists(clipseg_model_dir):
-    clipseg_model_dir='CIDAS/clipseg-rd64-refined'
-
-"""Helper methods for CLIPSeg nodes"""
-
-def tensor_to_numpy(tensor: torch.Tensor) -> np.ndarray:
-    """Convert a tensor to a numpy array and scale its values to 0-255."""
-    array = tensor.numpy().squeeze()
-    return (array * 255).astype(np.uint8)
-
-def numpy_to_tensor(array: np.ndarray) -> torch.Tensor:
-    """Convert a numpy array to a tensor and scale its values from 0-255 to 0-1."""
-    array = array.astype(np.float32) / 255.0
-    return torch.from_numpy(array)[None,]
-
-def apply_colormap(mask: torch.Tensor, colormap) -> np.ndarray:
-    """Apply a colormap to a tensor and convert it to a numpy array."""
-    colored_mask = colormap(mask.numpy())[:, :, :3]
-    return (colored_mask * 255).astype(np.uint8)
-
-def resize_image(image: np.ndarray, dimensions: Tuple[int, int]) -> np.ndarray:
-    """Resize an image to the given dimensions using linear interpolation."""
-    return cv2.resize(image, dimensions, interpolation=cv2.INTER_LINEAR)
-
-def overlay_image(background: np.ndarray, foreground: np.ndarray, alpha: float) -> np.ndarray:
-    """Overlay the foreground image onto the background with a given opacity (alpha)."""
-    return cv2.addWeighted(background, 1 - alpha, foreground, alpha, 0)
-
-def dilate_mask(mask: torch.Tensor, dilation_factor: float) -> torch.Tensor:
-    """Dilate a mask using a square kernel with a given dilation factor."""
-    kernel_size = int(dilation_factor * 2) + 1
-    kernel = np.ones((kernel_size, kernel_size), np.uint8)
-    mask_dilated = cv2.dilate(mask.numpy(), kernel, iterations=1)
-    return torch.from_numpy(mask_dilated)
-
-
-
-class CLIPSeg:
-
-    def __init__(self):
-        pass
-    
-    @classmethod
-    def INPUT_TYPES(s):
-        """
-            Return a dictionary which contains config for all input fields.
-            Some types (string): "MODEL", "VAE", "CLIP", "CONDITIONING", "LATENT", "IMAGE", "INT", "STRING", "FLOAT".
-            Input types "INT", "STRING" or "FLOAT" are special values for fields on the node.
-            The type can be a list for selection.
-
-            Returns: `dict`:
-                - Key input_fields_group (`string`): Can be either required, hidden or optional. A node class must have property `required`
-                - Value input_fields (`dict`): Contains input fields config:
-                    * Key field_name (`string`): Name of a entry-point method's argument
-                    * Value field_config (`tuple`):
-                        + First value is a string indicate the type of field or a list for selection.
-                        + Secound value is a config for type "INT", "STRING" or "FLOAT".
-        """
-        return {"required":
-                    {
-                        "image": ("IMAGE",),
-                        "text": ("STRING", {"multiline": False}),
-                        
-                     },
-                "optional":
-                    {
-                        "blur": ("FLOAT", {"min": 0, "max": 15, "step": 0.1, "default": 7}),
-                        "threshold": ("FLOAT", {"min": 0, "max": 1, "step": 0.05, "default": 0.4}),
-                        "dilation_factor": ("INT", {"min": 0, "max": 10, "step": 1, "default": 4}),
-                    }
-                }
-
-    CATEGORY = "♾️Mixlab/mask"
-    RETURN_TYPES = ("MASK", "IMAGE", "IMAGE",)
-    RETURN_NAMES = ("Mask","Heatmap Mask", "BW Mask")
-
-    # INPUT_IS_LIST = True
-    # OUTPUT_IS_LIST = (True,)
-
-    FUNCTION = "segment_image"
-    def segment_image(self, image: torch.Tensor, text: str, blur: float, threshold: float, dilation_factor: int) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-        """Create a segmentation mask from an image and a text prompt using CLIPSeg.
-
-        Args:
-            image (torch.Tensor): The image to segment.
-            text (str): The text prompt to use for segmentation.
-            blur (float): How much to blur the segmentation mask.
-            threshold (float): The threshold to use for binarizing the segmentation mask.
-            dilation_factor (int): How much to dilate the segmentation mask.
-
-        Returns:
-            Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: The segmentation mask, the heatmap mask, and the binarized mask.
-        """
-            
-        # Convert the Tensor to a PIL image
-        image_np = image.numpy().squeeze()  # Remove the first dimension (batch size of 1)
-        # Convert the numpy array back to the original range (0-255) and data type (uint8)
-        image_np = (image_np * 255).astype(np.uint8)
-        # Create a PIL image from the numpy array
-        i = Image.fromarray(image_np, mode="RGB")
-
-        processor = CLIPSegProcessor.from_pretrained(clipseg_model_dir)
-        model = CLIPSegForImageSegmentation.from_pretrained(clipseg_model_dir)
-        
-        prompt = text
-        
-        input_prc = processor(text=prompt, images=i, padding="max_length", return_tensors="pt")
-        
-        # Predict the segemntation mask
-        with torch.no_grad():
-            outputs = model(**input_prc)
-        
-        tensor = torch.sigmoid(outputs[0]) # get the mask
-        
-        # Apply a threshold to the original tensor to cut off low values
-        thresh = threshold
-        tensor_thresholded = torch.where(tensor > thresh, tensor, torch.tensor(0, dtype=torch.float))
-
-        # Apply Gaussian blur to the thresholded tensor
-        sigma = blur
-        tensor_smoothed = gaussian_filter(tensor_thresholded.numpy(), sigma=sigma)
-        tensor_smoothed = torch.from_numpy(tensor_smoothed)
-
-        # Normalize the smoothed tensor to [0, 1]
-        mask_normalized = (tensor_smoothed - tensor_smoothed.min()) / (tensor_smoothed.max() - tensor_smoothed.min())
-
-        # Dilate the normalized mask
-        mask_dilated = dilate_mask(mask_normalized, dilation_factor)
-
-        # Convert the mask to a heatmap and a binary mask
-        heatmap = apply_colormap(mask_dilated, cm.viridis)
-        binary_mask = apply_colormap(mask_dilated, cm.Greys_r)
-
-        # Overlay the heatmap and binary mask on the original image
-        dimensions = (image_np.shape[1], image_np.shape[0])
-        heatmap_resized = resize_image(heatmap, dimensions)
-        binary_mask_resized = resize_image(binary_mask, dimensions)
-
-        alpha_heatmap, alpha_binary = 0.5, 1
-        overlay_heatmap = overlay_image(image_np, heatmap_resized, alpha_heatmap)
-        overlay_binary = overlay_image(image_np, binary_mask_resized, alpha_binary)
-
-        # Convert the numpy arrays to tensors
-        image_out_heatmap = numpy_to_tensor(overlay_heatmap)
-        image_out_binary = numpy_to_tensor(overlay_binary)
-
-        # Save or display the resulting binary mask
-        binary_mask_image = Image.fromarray(binary_mask_resized[..., 0])
-
-        # convert PIL image to numpy array
-        tensor_bw = binary_mask_image.convert("RGB")
-        tensor_bw = np.array(tensor_bw).astype(np.float32) / 255.0
-        tensor_bw = torch.from_numpy(tensor_bw)[None,]
-        tensor_bw = tensor_bw.squeeze(0)[..., 0]
-
-        return tensor_bw, image_out_heatmap, image_out_binary
-
-    #OUTPUT_NODE = False
-
-class CombineMasks:
-    def __init__(self):
-        pass
-
-    @classmethod
-    def INPUT_TYPES(s):
-        return {"required":
-                    {
-                        "input_image": ("IMAGE", ),
-                        "mask_1": ("MASK", ), 
-                        "mask_2": ("MASK", ),
-                    },
-                "optional": 
-                    {
-                        "mask_3": ("MASK",), 
-                    },
-                }
-        
-    CATEGORY = "♾️Mixlab/mask"
-    RETURN_TYPES = ("MASK", "IMAGE", "IMAGE",)
-    RETURN_NAMES = ("Combined Mask","Heatmap Mask", "BW Mask")
-
-    FUNCTION = "combine_masks"
-            
-    def combine_masks(self, input_image: torch.Tensor, mask_1: torch.Tensor, mask_2: torch.Tensor, mask_3: Optional[torch.Tensor] = None) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-        """A method that combines two or three masks into one mask. Takes in tensors and returns the mask as a tensor, as well as the heatmap and binary mask as tensors."""
-        
-        # Combine masks
-        if mask_1 is not None:
-            mask_1 = mask_1.squeeze()
-        if mask_2 is not None:
-            mask_2 = mask_2.squeeze()
-        if mask_3 is not None:
-            mask_3 = mask_3.squeeze()
-        
-        print(mask_1.shape,mask_2.shape , mask_3.shape)
-        combined_mask = mask_1 + mask_2 + mask_3 if mask_3 is not None else mask_1 + mask_2
-        # print(combined_mask)
-
-        # Convert image and masks to numpy arrays
-        image_np = tensor_to_numpy(input_image)
-        heatmap = apply_colormap(combined_mask, cm.viridis)
-        binary_mask = apply_colormap(combined_mask, cm.Greys_r)
-        
-        # Resize heatmap and binary mask to match the original image dimensions
-        dimensions = (image_np.shape[1], image_np.shape[0])
-        print('heatmap',heatmap)
-        if dimensions is None or dimensions[0] == 0 or dimensions[1] == 0:
-            raise ValueError("Invalid dimensions")
-        
-        heatmap_resized = resize_image(heatmap, dimensions)
-        binary_mask_resized = resize_image(binary_mask, dimensions)
-
-        # Overlay the heatmap and binary mask onto the original image
-        alpha_heatmap, alpha_binary = 0.5, 1
-        overlay_heatmap = overlay_image(image_np, heatmap_resized, alpha_heatmap)
-        overlay_binary = overlay_image(image_np, binary_mask_resized, alpha_binary)
-
-        # Convert overlays to tensors
-        image_out_heatmap = numpy_to_tensor(overlay_heatmap)
-        image_out_binary = numpy_to_tensor(overlay_binary)
-
-        return combined_mask, image_out_heatmap, image_out_binary
-
-# A dictionary that contains all nodes you want to export with their names
-# NOTE: names should be globally unique
-# NODE_CLASS_MAPPINGS = {
-#     "CLIPSeg": CLIPSeg,
-#     "CombineSegMasks": CombineMasks,
-# }
@@ -0,0 +1,123 @@
+import os,sys
+import folder_paths
+
+from PIL import Image
+import importlib.util
+import numpy as np
+
+import torch
+
+global _available
+_available=False
+
+def is_installed(package):
+    try:
+        spec = importlib.util.find_spec(package)
+    except ModuleNotFoundError:
+        return False
+    return spec is not None
+
+if is_installed('simple_lama_inpainting')==False:
+    import subprocess
+    from packaging import version
+    
+    if version.parse(torch.__version__)>=version.parse('2.1'):
+        # 安装
+        print('#pip install simple_lama_inpainting')
+
+        result = subprocess.run([sys.executable, '-s', '-m', 'pip', 'install', 'simple_lama_inpainting'], capture_output=True, text=True)
+
+        #检查命令执行结果
+        if result.returncode == 0:
+            print("#install success")
+            from simple_lama_inpainting import SimpleLama
+            _available=True
+        else:
+            print("#install error")
+    else:
+        print('#pls check your torch version >= 2.1')
+
+else:
+    from simple_lama_inpainting import SimpleLama
+    _available=True
+
+
+def get_lama_path():
+    try:
+        return folder_paths.get_folder_paths('lama')[0]
+    except:
+        return os.path.join(folder_paths.models_dir, "lama")
+
+llma_model_path=os.path.join(get_lama_path(), "big-lama.pt")
+if not os.path.exists(llma_model_path):
+    os.environ['LAMA_MODEL']=''
+    print(f"## lama torchscript model not found: {llma_model_path},pls download from https://github.com/enesmsahin/simple-lama-inpainting/releases/download/v0.1.0/big-lama.pt")
+else:
+    os.environ['LAMA_MODEL'] = llma_model_path
+
+# Tensor to PIL
+def tensor2pil(image):
+    return Image.fromarray(np.clip(255. * image.cpu().numpy().squeeze(), 0, 255).astype(np.uint8))
+
+# Convert PIL to Tensor
+def pil2tensor(image):
+    return torch.from_numpy(np.array(image).astype(np.float32) / 255.0).unsqueeze(0)
+
+
+# simple_lama = SimpleLama()
+
+# img_path = "image.png"
+# mask_path = "mask.png"
+
+# image = Image.open(img_path)
+# mask = Image.open(mask_path).convert('L')
+
+# result = simple_lama(image, mask)
+# result.save("inpainted.png")
+
+
+class LaMaInpainting:
+    global _available
+    available=_available
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "image": ("IMAGE",),
+            "mask": ("MASK",),
+                             },
+                }
+    
+    RETURN_TYPES = ("IMAGE",)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Image"
+
+    INPUT_IS_LIST = True
+    OUTPUT_IS_LIST = (True,)
+    global simple_lama
+    simple_lama = None
+    def run(self,image,mask):
+        global simple_lama
+
+        result=[]
+        if simple_lama==None:
+            simple_lama = SimpleLama()
+        else:
+            simple_lama.model.to("cuda" if torch.cuda.is_available() else "cpu")
+
+        for i in range(len(image)):
+            im=image[i]
+            ma=mask[i]
+            im=tensor2pil(im)
+            ma=tensor2pil(ma)
+            ma =ma.convert('L')
+            
+            res = simple_lama(im, ma)
+            res=pil2tensor(res)
+            result.append(res)
+        # result.save("inpainted.png")
+        if simple_lama.device=='cuda':
+            simple_lama.model.to('cpu')
+
+        return (result,)
@@ -0,0 +1,274 @@
+ 
+import scipy.ndimage
+import torch
+
+import numpy as np 
+# from PIL import Image, ImageDraw
+from PIL import Image, ImageOps 
+
+from comfy.cli_args import args
+import cv2,os
+from nodes import MAX_RESOLUTION, SaveImage, common_ksampler
+import folder_paths,random
+
+# Tensor to PIL
+def tensor2pil(image):
+    return Image.fromarray(np.clip(255. * image.cpu().numpy().squeeze(), 0, 255).astype(np.uint8))
+
+# Convert PIL to Tensor
+def pil2tensor(image):
+    return torch.from_numpy(np.array(image).astype(np.float32) / 255.0).unsqueeze(0)
+
+
+def add_masks(mask1, mask2):
+    mask1 = mask1.cpu()
+    mask2 = mask2.cpu()
+    cv2_mask1 = np.array(mask1) * 255
+    cv2_mask2 = np.array(mask2) * 255
+
+    if cv2_mask1.shape == cv2_mask2.shape:
+        cv2_mask = cv2.add(cv2_mask1, cv2_mask2)
+        return torch.clamp(torch.from_numpy(cv2_mask) / 255.0, min=0, max=1)
+    else:
+        return mask1
+    
+
+def grow(mask, expand, tapered_corners):
+    c = 0 if tapered_corners else 1
+    kernel = np.array([[c, 1, c],
+                            [1, 1, 1],
+                            [c, 1, c]])
+    mask = mask.reshape((-1, mask.shape[-2], mask.shape[-1]))
+    out = []
+    for m in mask:
+        output = m.numpy()
+        for _ in range(abs(expand)):
+            if expand < 0:
+                output = scipy.ndimage.grey_erosion(output, footprint=kernel)
+            else:
+                output = scipy.ndimage.grey_dilation(output, footprint=kernel)
+        output = torch.from_numpy(output)
+        out.append(output)
+    return torch.stack(out, dim=0)
+
+def combine(destination, source, x, y):
+    output = destination.reshape((-1, destination.shape[-2], destination.shape[-1])).clone()
+    source = source.reshape((-1, source.shape[-2], source.shape[-1]))
+
+    left, top = (x, y,)
+    right, bottom = (min(left + source.shape[-1], destination.shape[-1]), min(top + source.shape[-2], destination.shape[-2]))
+    visible_width, visible_height = (right - left, bottom - top,)
+
+    source_portion = source[:, :visible_height, :visible_width]
+    destination_portion = destination[:, top:bottom, left:right]
+ 
+    #operation == "subtract":
+    output[:, top:bottom, left:right] = destination_portion - source_portion
+        
+    output = torch.clamp(output, 0.0, 1.0)
+
+    return output
+
+
+class PreviewMask_(SaveImage):
+    def __init__(self):
+        self.output_dir = folder_paths.get_temp_directory()
+        self.type = "temp"
+        self.prefix_append =''.join(random.choice("abcdehijklmnopqrstupvxyzfg") for x in range(5))
+        self.compress_level = 4
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+                "required": {
+                                "mask": ("MASK",), 
+                            }
+            }
+    
+    RETURN_TYPES = ()
+    OUTPUT_NODE = True
+    FUNCTION = "run"
+    CATEGORY = "♾️Mixlab/Mask"
+
+    # 运行的函数
+    def run(self, mask ):
+        img=tensor2pil(mask)
+        img=img.convert('RGB')
+        img=pil2tensor(img) 
+        return self.save_images(img, 'temp_', None,  None)
+
+
+class OutlineMask:
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+                "required": {
+                                "mask": ("MASK",),
+                                "outline_width":("INT", {"default": 10,"min": 1, "max": MAX_RESOLUTION, "step": 1}),
+                                "tapered_corners": ("BOOLEAN", {"default": True}),
+                            }
+            }
+    
+    RETURN_TYPES = ('MASK',)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Mask"
+
+    # 运行的函数
+    def run(self, mask, outline_width, tapered_corners):
+
+        m1=grow(mask,outline_width,tapered_corners)
+        m2=grow(mask,-outline_width,tapered_corners)
+
+        m3=combine(m1,m2,0,0)
+
+        return (m3,)
+    
+
+class MaskListReplace:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "masks": ("MASK",),
+            "mask_replace": ("MASK",),
+            "start_index":("INT", {"default": 0, "min": 0, "step": 1}),
+            "end_index":("INT", {"default": 0, "min": 0, "step": 1}),
+            "invert": ("BOOLEAN", {"default": False}),
+        }
+        }
+
+    RETURN_TYPES = ("MASK",)
+    FUNCTION = "run"
+    CATEGORY = "♾️Mixlab/Video"
+
+    INPUT_IS_LIST = True
+    OUTPUT_IS_LIST = (True,)
+
+    def run(self, masks,mask_replace,start_index,end_index,invert):
+        mask_replace=mask_replace[0]
+        start_index=start_index[0]
+        end_index=end_index[0]
+        invert=invert[0]
+        
+        new_masks=[]
+        for i in range(len(masks)):
+            if i>=start_index and i<=end_index:
+                if invert:
+                    new_masks.append(masks[i])
+                else:
+                    new_masks.append(mask_replace)
+            else:
+                if invert:
+                    new_masks.append(mask_replace)
+                else:
+                    new_masks.append(masks[i])
+
+        return (new_masks,)
+
+
+class MaskListMerge:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "masks": ("MASK",),
+        }
+        }
+
+    RETURN_TYPES = ("MASK",)
+    FUNCTION = "run"
+    CATEGORY = "♾️Mixlab/Mask"
+
+    INPUT_IS_LIST = True
+    OUTPUT_IS_LIST = (False,)
+
+    def run(self, masks):
+        mask=masks[0]
+        if isinstance(masks, list):
+            for m in masks:
+                # print(m.shape)
+                mask = add_masks(mask, m)
+        return (mask,)
+
+
+class FeatheredMask:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+                "required": {
+                                "mask": ("MASK",),
+                                "start_offset":("INT", {"default": 1, 
+                                                        "min": -150, 
+                                                        "max": 150, 
+                                                        "step": 1,
+                                                        "display": "slider"}),
+                                "feathering_weight":("FLOAT", {"default": 0.1,
+                                                                "min": 0.0,
+                                                                "max": 1,
+                                                                "step": 0.1,
+                                                                "display": "slider"})
+                            }
+            }
+    
+    RETURN_TYPES = ('MASK',)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Mask"
+
+    OUTPUT_IS_LIST = (True,)
+  
+    # 运行的函数
+    def run(self,mask,start_offset, feathering_weight):
+        # print(mask.shape,mask.size())
+        
+        num,_,_=mask.size()
+
+        masks=[]
+
+        for i in range(num):
+            mm=mask[i]
+            image=tensor2pil(mm)
+
+            # Open the image using PIL
+            image = image.convert("L")
+            if start_offset>0:
+                image=ImageOps.invert(image)
+
+            # Convert the image to a numpy array
+            image_np = np.array(image)
+
+            # Use Canny edge detection to get black contours
+            edges = cv2.Canny(image_np, 30, 150)
+
+            for i in range(0,abs(start_offset)):
+                # int(100*feathering_weight)
+                a=int(abs(start_offset)*0.1*i)
+                # Dilate the black contours to make them wider
+                kernel = np.ones((a, a), np.uint8)
+
+                dilated_edges = cv2.dilate(edges, kernel, iterations=1)
+                # dilated_edges = cv2.erode(edges, kernel, iterations=1)
+                # Smooth the dilated edges using Gaussian blur
+                smoothed_edges = cv2.GaussianBlur(dilated_edges, (5, 5), 0)
+
+                # Adjust the feathering weight
+                feathering_weight = max(0, min(feathering_weight, 1))
+
+                # Blend the smoothed edges with the original image to achieve feathering effect
+                image_np = cv2.addWeighted(image_np, 1, smoothed_edges, feathering_weight, feathering_weight)
+
+            # Convert the result back to PIL image
+            result_image = Image.fromarray(np.uint8(image_np))
+            result_image=result_image.convert("L")
+
+            if start_offset>0:
+                result_image=ImageOps.invert(result_image)
+            
+            result_image=result_image.convert("L")
+            mt=pil2tensor(result_image)
+            masks.append(mt)
+         
+            # print( mt.size())
+        return (masks,)
@@ -0,0 +1,127 @@
+# Referenced some code：https://github.com/IuvenisSapiens/ComfyUI_MiniCPM-V-2_6-int4
+
+import os
+import torch
+import folder_paths
+from transformers import AutoTokenizer, AutoModel
+from torchvision.transforms.v2 import ToPILImage
+# from decord import VideoReader, cpu  # pip install decord
+from PIL import Image
+
+def get_model_path(n=""):
+    try:
+        return folder_paths.get_folder_paths(n)[0]
+    except:
+        return os.path.join(folder_paths.models_dir, n)
+ 
+
+class MiniCPM_VQA_Simple:
+    def __init__(self):
+        self.model_checkpoint = None
+        self.tokenizer = None
+        self.model = None
+        self.device = (
+            torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+        )
+        self.bf16_support = (
+            torch.cuda.is_available()
+            and torch.cuda.get_device_capability(self.device)[0] >= 8
+        )
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "images": ("IMAGE",),
+                "text": ("STRING", {"default": "", "multiline": True}),   
+                "seed": ("INT", {"default": -1}),  # add seed parameter, default is -1
+                "temperature": (
+                    "FLOAT",
+                    {
+                        "default": 0.7,
+                    },
+                ),
+                "keep_model_loaded": ("BOOLEAN", {"default": False}),
+            },
+            
+        }
+
+    RETURN_TYPES = ("STRING",)
+    FUNCTION = "inference"
+    CATEGORY = "♾️Mixlab/Image"
+
+    def inference(
+        self,
+        images,
+        text, 
+        seed,  # add seed parameter, default is -1
+        temperature,
+        keep_model_loaded,
+    ):
+        if seed != -1:
+            torch.manual_seed(seed)
+        model_id = "openbmb/MiniCPM-V-2_6-int4"
+
+        self.model_checkpoint = os.path.join( get_model_path("prompt_generator"), os.path.basename(model_id))
+
+        if not os.path.exists(self.model_checkpoint):
+            from huggingface_hub import snapshot_download
+
+            snapshot_download(
+                repo_id=model_id,
+                local_dir=self.model_checkpoint,
+                local_dir_use_symlinks=False,
+                endpoint='https://hf-mirror.com'
+            )
+
+        if self.tokenizer is None:
+            self.tokenizer = AutoTokenizer.from_pretrained(
+                self.model_checkpoint,
+                trust_remote_code=True,
+                low_cpu_mem_usage=True,
+            )
+
+        if self.model is None:
+            self.model = AutoModel.from_pretrained(
+                self.model_checkpoint,
+                trust_remote_code=True,
+                low_cpu_mem_usage=True,
+                attn_implementation="sdpa",
+                torch_dtype=torch.bfloat16 if self.bf16_support else torch.float16,
+            )
+
+        with torch.no_grad():
+            images = images.permute([0, 3, 1, 2])
+            images = [ToPILImage()(img).convert("RGB") for img in images]
+            msgs = [{"role": "user", "content": images + [text]}]
+
+            params = {"use_image_id": False, }
+
+            # offload model to CPU
+            # self.model = self.model.to(torch.device("cpu"))
+            # self.model.eval()
+
+            result = self.model.chat(
+                image=None,
+                msgs=msgs,
+                tokenizer=self.tokenizer,
+                sampling=True,
+                # top_k=top_k,
+                # top_p=top_p,
+                temperature=temperature,
+                # repetition_penalty=repetition_penalty,
+                # max_new_tokens=max_new_tokens,
+                **params,
+            )
+            # offload model to GPU
+            # self.model = self.model.to(torch.device("cpu"))
+            # self.model.eval()
+            if not keep_model_loaded:
+                del self.tokenizer  # release tokenizer memory
+                del self.model  # release model memory
+                self.tokenizer = None  # set tokenizer to None
+                self.model = None  # set model to None
+                torch.cuda.empty_cache()  # release GPU memory
+                torch.cuda.ipc_collect()
+
+            return (result,)
@@ -0,0 +1,104 @@
+import torch
+import numpy as np
+from PIL import Image,ImageSequence,ImageOps
+import base64
+import io
+import comfy.utils
+import folder_paths
+import node_helpers
+
+
+# Tensor to PIL
+def tensor2pil(image):
+    return Image.fromarray(np.clip(255. * image.cpu().numpy().squeeze(), 0, 255).astype(np.uint8))
+
+# Convert PIL to Tensor
+def pil2tensor(image):
+    return torch.from_numpy(np.array(image).astype(np.float32) / 255.0).unsqueeze(0)
+
+def load_image_to_tensor( image):
+    image_path = folder_paths.get_annotated_filepath(image)
+        
+    img = node_helpers.pillow(Image.open, image_path)
+        
+    output_images = []
+    output_masks = []
+    w, h = None, None
+
+    excluded_formats = ['MPO']
+        
+    for i in ImageSequence.Iterator(img):
+        i = node_helpers.pillow(ImageOps.exif_transpose, i)
+
+        if i.mode == 'I':
+            i = i.point(lambda i: i * (1 / 255))
+        image = i.convert("RGB")
+
+        if len(output_images) == 0:
+            w = image.size[0]
+            h = image.size[1]
+            
+        if image.size[0] != w or image.size[1] != h:
+            continue
+            
+        image = np.array(image).astype(np.float32) / 255.0
+        image = torch.from_numpy(image)[None,]
+        if 'A' in i.getbands():
+            mask = np.array(i.getchannel('A')).astype(np.float32) / 255.0
+            mask = 1. - torch.from_numpy(mask)
+        else:
+            mask = torch.zeros((64,64), dtype=torch.float32, device="cpu")
+        output_images.append(image)
+        output_masks.append(mask.unsqueeze(0))
+
+    if len(output_images) > 1 and img.format not in excluded_formats:
+        output_image = torch.cat(output_images, dim=0)
+        output_mask = torch.cat(output_masks, dim=0)
+    else:
+        output_image = output_images[0]
+        output_mask = output_masks[0]
+
+    return (output_image, output_mask)
+
+          
+
+class P5Input:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { 
+                    "frames":("IMAGEBASE64",), 
+                             },
+                }
+    
+    RETURN_TYPES = ("IMAGE",)
+    RETURN_NAMES = ("frames",)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Input"
+
+    OUTPUT_NODE = True
+    INPUT_IS_LIST = False
+    OUTPUT_IS_LIST = (False,)
+
+    def run(self, frames):
+        ims=[]
+        for im in frames['images']:
+            # print(im)
+            if 'type' in im and (not f"[{im['type']}]" in im['name']):
+                im['name']=im['name']+" "+f"[{im['type']}]"
+                
+            output_image, output_mask = load_image_to_tensor(im['name'])
+            ims.append(output_image)
+
+        if len(ims)==0:
+            image1 = Image.new('RGB', (512, 512), color='black')
+            return (pil2tensor(image1),)
+        image1 = ims[0]
+        for image2 in ims[1:]:
+            if image1.shape[1:] != image2.shape[1:]:
+                image2 = comfy.utils.common_upscale(image2.movedim(-1, 1), image1.shape[2], image1.shape[1], "bilinear", "center").movedim(1, -1)
+            image1 = torch.cat((image1, image2), dim=0)
+
+        # 用于节点提示:p5节点提示有多少帧
+        return {"ui": {"_info": [len(frames['images'])]}, "result": (image1,)}
@@ -1,15 +1,97 @@
 import random
 import comfy.utils
-import json
+import os
+import numpy as np
 from urllib import request, parse
+import folder_paths
+from PIL import Image, ImageOps,ImageFilter,ImageEnhance,ImageDraw,ImageSequence, ImageFont
+from PIL.PngImagePlugin import PngInfo
+
+import hashlib
+import requests
+import json
+
+
+# def queue_prompt(prompt_workflow):
+#     p = {"prompt": prompt_workflow}
+#     data = json.dumps(p).encode('utf-8')
+#     req =  request.Request("http://127.0.0.1:8188/prompt", data=data)
+#     request.urlopen(req)    
+
+def get_model_path(n=""):
+    try:
+        return folder_paths.get_folder_paths(n)[0]
+    except:
+        return os.path.join(folder_paths.models_dir, n)
+    
+embeddings_path=get_model_path("embeddings")
+
+def get_files_with_extension(directory, extension):
+    
+    file_list = []
+    for root, dirs, files in os.walk(directory):
+        for file in files:
+            if file.endswith(extension):
+                file_name = os.path.splitext(file)[0]
+                file_list.append(file_name)
+    return file_list
+
+def join_with_(text_list,delimiter):
+    joined_text = delimiter.join(text_list)
+    return joined_text



-def queue_prompt(prompt_workflow):
-    p = {"prompt": prompt_workflow}
-    data = json.dumps(p).encode('utf-8')
-    req =  request.Request("http://127.0.0.1:8188/prompt", data=data)
-    request.urlopen(req)    
+def load_json(file_path):
+    try:
+        with open(file_path, 'r') as json_file:
+            data = json.load(json_file)
+            return data
+    except FileNotFoundError:
+        print(f"File not found: {file_path}")
+        return None
+    except json.JSONDecodeError:
+        print(f"Error decoding JSON in file: {file_path}")
+        return None
+
+def save_json(data_dict, file_path):
+    try:
+        with open(file_path, 'w') as json_file:
+            json.dump(data_dict, json_file, indent=4)
+            print(f"Data saved to {file_path}")
+    except Exception as e:
+        print(f"Error saving JSON to file: {e}")
+
+# pysss的lora加载器
+# def get_model_version_info(hash_value):
+#     # http://127.0.0.1:1082
+#     proxies = {'http': 'http://127.0.0.1:1082', 'https': 'https://127.0.0.1:1082'}
+#     api_url = f"https://civitai.com/api/v1/model-versions/by-hash/{hash_value}"
+#     print(api_url)
+#     response = requests.get(api_url,proxies=proxies, verify=False)
+    
+#     if response.status_code == 200:
+#         return response.json()
+#     else:
+#         return None
+    
+# def calculate_sha256(file_path):
+#     sha256_hash = hashlib.sha256()
+#     with open(file_path, "rb") as f:
+#         for chunk in iter(lambda: f.read(4096), b""):
+#             sha256_hash.update(chunk)
+#     return sha256_hash.hexdigest()
+
+
+
+
+class AnyType(str):
+  """A special class that is always equal in not equal comparisons. Credit to pythongosssss"""
+
+  def __ne__(self, __value: object) -> bool:
+    return False
+
+any_type = AnyType("*")


 default_prompt1='''Swing
@@ -45,6 +127,232 @@ default_prompt1='''Swing
                                                '''
 default_prompt1="\n".join([p.strip() for p in default_prompt1.split('\n') if p.strip()!=''])

+
+def tensor2pil(image):
+    return Image.fromarray(np.clip(255. * image.cpu().numpy().squeeze(), 0, 255).astype(np.uint8))
+
+def addWeight(text, weight=1):
+    if weight == 1:
+        return text
+    else:
+        return f"({text}:{round(weight,3)})"
+
+def prompt_delete_words(sentence, new_words_length):
+    # 使用逗号分割句子，并去除空格
+    words = [word.strip() for word in sentence.split(",")]
+    
+    # 计算需要删除的单词数量
+    num_to_delete = len(words) - new_words_length
+    
+    words_to=[w for w in words]
+
+    # 逐个删除单词并存储在新列表中
+    new_words = []
+    for i in range(len(words)):
+        if num_to_delete > 0:
+            num_to_delete -= 1
+        else:
+            words_to.pop()
+            if len(words_to)>0:
+                new_words.append(", ".join(words_to))
+         
+    return new_words
+
+# # 测试方法
+# sentence = "a computer, a glass tablet with a keyboard on a dark background, 3d illustration, reflection, cgi 8k, clear glass, archaic, cut-away, white outline"
+# new_words_length = 5
+# result = prompt_delete_words(sentence, new_words_length)
+# print(result)
+
+class PromptImage:
+    def __init__(self):
+        self.output_dir = folder_paths.get_output_directory()
+        self.type = "output"
+        self.prefix_append = "PromptImage"
+        self.compress_level = 4
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "prompts": ("STRING", 
+                         {
+                            "multiline": True, 
+                            "default": '',
+                            "dynamicPrompts": False
+                          }),
+
+                "images": ("IMAGE",{"default": None}), 
+                 "save_to_image": (["enable", "disable"],),
+                }
+            }
+    
+    RETURN_TYPES = ()
+   
+    OUTPUT_NODE = True
+
+    INPUT_IS_LIST = True
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Output"
+
+    # 运行的函数
+    def run(self,prompts,images,save_to_image):
+        filename_prefix="mixlab_"
+        filename_prefix += self.prefix_append
+        full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(
+            filename_prefix, self.output_dir, images[0].shape[1], images[0].shape[0])
+        
+        results = list()
+
+        save_to_image=save_to_image[0]=='enable'
+
+        for index in range(len(images)):
+            res=[]
+            imgs=images[index]
+
+            for image in imgs:
+                img=tensor2pil(image)
+
+                metadata = None
+                if save_to_image:
+                    metadata = PngInfo()
+                    prompt_text=prompts[index]
+                    if prompt_text is not None:
+                        metadata.add_text("prompt_text", prompt_text)
+                    
+                file = f"{filename}_{index}_{counter:05}_.png"
+                img.save(os.path.join(full_output_folder, file), pnginfo=metadata, compress_level=self.compress_level)
+                res.append({
+                    "filename": file,
+                    "subfolder": subfolder,
+                    "type": self.type
+                })
+                counter += 1
+            results.append(res)
+        
+        return { "ui": { "_images": results,"prompts":prompts } }
+
+
+
+
+class PromptSimplification:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "prompt": ("STRING", 
+                         {
+                            "multiline": True, 
+                            "default": '',
+                            "dynamicPrompts": False
+                          }),
+
+                "length":("INT", {"default": 5, "min": 1,"max":100, "step": 1, "display": "number"}),
+
+                # "min_value":("FLOAT", {
+                #         "default": -2, 
+                #         "min": -10, 
+                #         "max": 0xffffffffffffffff,
+                #         "step": 0.01, 
+                #         "display": "number"  
+                #     }),
+                # "max_value":("FLOAT", {
+                #         "default": 2, 
+                #         "min": -10, 
+                #         "max": 0xffffffffffffffff,
+                #         "step": 0.01, 
+                #         "display": "number"  
+                #     }),
+              
+                }
+            }
+    
+    RETURN_TYPES = ("STRING",)
+    RETURN_NAMES = ("prompts",)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Prompt"
+
+    INPUT_IS_LIST = True
+    OUTPUT_IS_LIST = (True,)
+    OUTPUT_NODE = True
+
+    # 运行的函数
+    def run(self,prompt,length):
+        length=length[0]
+        result=[]
+        for p in prompt:
+            nps=prompt_delete_words(p,length)
+            for n in nps:
+                result.append(n)
+
+        result= [elem.strip() for elem in result if elem.strip()]
+
+        return {"ui": {"prompts": result}, "result": (result,)}
+
+
+
+class PromptSlide:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                
+                "prompt_keyword": ("STRING", 
+                         {
+                            "multiline": False, 
+                            "default": '',
+                            "dynamicPrompts": False
+                          }),
+
+                "weight":("FLOAT", {"default": 1, "min": -3,"max": 3,
+                                                                "step": 0.01,
+                                                                "display": "slider"}),
+
+                # "min_value":("FLOAT", {
+                #         "default": -2, 
+                #         "min": -10, 
+                #         "max": 0xffffffffffffffff,
+                #         "step": 0.01, 
+                #         "display": "number"  
+                #     }),
+                # "max_value":("FLOAT", {
+                #         "default": 2, 
+                #         "min": -10, 
+                #         "max": 0xffffffffffffffff,
+                #         "step": 0.01, 
+                #         "display": "number"  
+                #     }),
+              
+                }
+            }
+    
+    RETURN_TYPES = ("STRING",)
+    RETURN_NAMES = ("prompt",)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Prompt"
+
+    INPUT_IS_LIST = False
+    OUTPUT_IS_LIST = (False,)
+    OUTPUT_NODE = False
+
+    # 运行的函数
+    def run(self,prompt_keyword,weight):
+        # if weight < min_value:
+        #     weight= min_value
+        # elif weight > max_value:
+        #     weight= max_value
+        p=addWeight(prompt_keyword,weight)
+        return (p,)
+
+
+
+
 class RandomPrompt:

    '''
@@ -70,6 +378,10 @@ class RandomPrompt:
                            "default": 'sticker, Cartoon, ``'
                          }),
                "random_sample": (["enable", "disable"],),
+                # "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff, "step": 1}),
+                },
+            "optional":{
+                    "seed": (any_type,  {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
                }
            }
    
@@ -79,15 +391,15 @@ class RandomPrompt:

    FUNCTION = "run"

-    CATEGORY = "♾️Mixlab/prompt"
+    CATEGORY = "♾️Mixlab/Prompt"

    OUTPUT_IS_LIST = (True,)
    OUTPUT_NODE = True


    # 运行的函数
-    def run(self,max_count,mutable_prompt,immutable_prompt,random_sample):
-        print('#运行的函数',mutable_prompt,immutable_prompt,max_count,random_sample)
+    def run(self,max_count,mutable_prompt,immutable_prompt,random_sample,seed=0):
+        # print('#运行的函数',mutable_prompt,immutable_prompt,max_count,random_sample)
        
        # Split the text into an array of words
        words1 = mutable_prompt.split("\n")
@@ -106,6 +418,11 @@ class RandomPrompt:
            w1=w1.strip()
            for w2 in words2:
                w2=w2.strip()
+                if '``' not in w2:
+                    if w2=="":
+                        w2='``'
+                    else:
+                        w2=w2+',``'
                if w1!='' and w2!='':
                    prompts.append(w2.replace('``', w1))
                pbar.update(1)
@@ -119,69 +436,258 @@ class RandomPrompt:
        else:
            prompts = prompts[:min(max_count,len(prompts))]

+        prompts= [elem.strip() for elem in prompts if elem.strip()]
+
        # return (new_prompt)
        return {"ui": {"prompts": prompts}, "result": (prompts,)}


+# class LoraPrompt:
+#     @classmethod
+#     def INPUT_TYPES(s):
+#         return {
+#             "required": {
+#                 "lora_name":(sorted(folder_paths.get_filename_list("loras"), key=str.lower),),
+#                 "weight": ("FLOAT", {"default": 1, "min": -2, "max": 2,"step":0.01 ,"display": "slider"}),
+#                 "force_update": ("BOOLEAN", {"default": False}),
+#                 },
+                
+#             }
+    
+#     RETURN_TYPES = ("STRING","STRING",any_type)
+#     RETURN_NAMES = ("lora_name","prompt","tags",)
+
+#     FUNCTION = "run"
+
+#     CATEGORY = "♾️Mixlab/Prompt"
+
+#     OUTPUT_IS_LIST = (False,False,True,)
+#     # OUTPUT_NODE = True
+
+#     # 运行的函数
+#     def run(self,lora_name,weight,force_update=False):
+       
+#         # print('##LoraPrompt',__file__)
+#         # 从本地数据库读取
+#         json_tags_path = os.path.join(os.path.dirname(os.path.dirname(__file__)),r'data/loras_tags.json')
+
+#         if not os.path.exists(json_tags_path):
+#             save_json({},json_tags_path)
+
+#         lora_tags = load_json(json_tags_path)
+#         output_tags = lora_tags.get(lora_name, None) if lora_tags is not None else None
+#         if output_tags is not None:
+#             output_tags = ",".join(output_tags)
+#             print("trainedWords:",output_tags)
+#         else:
+#             output_tags = ""
+
+
+#         lora_path = folder_paths.get_full_path("loras", lora_name)
+#         if output_tags == "" or force_update:
+#             print("calculating lora hash")
+#             LORAsha256 = calculate_sha256(lora_path)
+#             print("requesting infos")
+#             model_info = get_model_version_info(LORAsha256)
+#             if model_info is not None:
+#                 if "trainedWords" in model_info:
+#                     print("tags found!")
+#                     if lora_tags is None:
+#                         lora_tags = {}
+#                     lora_tags[lora_name] = model_info["trainedWords"]
+#                     save_json(lora_tags,json_tags_path)
+#                     output_tags = ",".join(model_info["trainedWords"])
+#                     print("trainedWords:",output_tags)
+#             else:
+#                 print("No informations found.")
+#                 if lora_tags is None:
+#                     lora_tags = {}
+#                 lora_tags[lora_name] = []
+#                 save_json(lora_tags,json_tags_path)
+
+
+#         weight = round(weight, 3)
+#         prompt=[]
+#         for p in output_tags.split(','):
+            
+#             if weight!=1:
+#                 prompt.append('('+p+':'+str(weight)+')')
+#             else:
+#                 prompt.append(p)
+
+#         prompt=",".join(prompt)
+
+#         return (lora_name,prompt,output_tags.split(','),)



-class RunWorkflow:
+class EmbeddingPrompt:
    @classmethod
    def INPUT_TYPES(s):
+
        return {
            "required": {
-                "workflow": ("STRING", {
-                            "multiline": False, 
-                            "default": ''
-                          }),
-                "prompt": ("STRING", {
-                            "multiline": False, 
-                            "default": ''
-                          }),
-                 "image": ("IMAGE",),
-                 "input_node": ("STRING", {
-                            "multiline": False, 
-                            "default": ''
-                          }),
-                 "output_node": ("STRING", {
-                            "multiline": False, 
-                            "default": ''
-                          }),
+                "embedding":(folder_paths.get_filename_list("embeddings"),),
+                "weight": ("FLOAT", {"default": 1, "min": -2, "max": 2,"step":0.01 ,"display": "slider"}),
                },
                
            }
    
-    
-
-    RETURN_TYPES = ("IMAGE","STRING",)
+    RETURN_TYPES = ("STRING",)

    FUNCTION = "run"

-    CATEGORY = "♾️Mixlab/workflow"
-
-    OUTPUT_IS_LIST = (True,)
-    OUTPUT_NODE = True
+    CATEGORY = "♾️Mixlab/Prompt"

+    OUTPUT_IS_LIST = (False,)
+    # OUTPUT_NODE = True

    # 运行的函数
-    def run(self,workflow,prompt,image,input_node,output_node):
-        print('#运行的函数',prompt,image,input_node,output_node)
-        workflow=json.loads(workflow)
-        input_node=input_node.split(".")
-        workflow[input_node[0]][input_node[1]][input_node[2]]=prompt
-
-        workflow_new={}
-        # 遍历，seed设为随机
-        for key, value in workflow.items():
-            if 'inputs' in value:
-                if 'seed' in value['inputs']:
-                    value['inputs']['seed']= random.randint(1, 18446744073709551614)
-            workflow_new[key]=value
-
-        queue_prompt(workflow_new)
-        print('#运行的函数',workflow_new[input_node[0]])
-
+    def run(self,embedding,weight):
+        weight = round(weight, 3)
+        prompt='embedding:'+embedding
+        if weight!=1:
+            prompt='('+prompt+':'+str(weight)+')'
+        prompt=" "+prompt+' ' 
        # return (new_prompt)
-        return  {"ui":{"images": []},"result": ([image],['text'],)}
+        return (prompt,)

+# RETURN_TYPES = (any_type,)
+    
+# conditioning ：提示，正向or负向
+# clip：clip模型
+# gligen_textbox_model：gligen模型
+# grids：矩形框的集合
+# labels：每个矩形框对应的标签的集合
+# index：选取第几个矩形框作为gligen的box
+
+class GLIGENTextBoxApply_Advanced:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"conditioning": ("CONDITIONING", ),
+                              "clip": ("CLIP", ),
+                              "gligen_textbox_model": ("GLIGEN", ),
+                              "grids": ("_GRID",),
+                              "labels": ("STRING", 
+                                        {
+                                            "multiline": True, 
+                                            "default": "",
+                                            "forceInput": True
+                                        }),
+                              "index": ("INT", {"default": -1, "min": -1, "max": 300, "step": 1}),
+                              "max_size": ("INT", {"default": 8, "min": 1, "max": 300, "step": 1}),
+                              "random_shuffle":(["on","off"],),
+                             },
+                 "optional":{
+                    "seed": (any_type,  {"default": 0, "min": 0, "max": 0xffffffffffffffff,"step": 1}),
+                } 
+                            }
+    RETURN_TYPES = ("CONDITIONING","STRING",)
+    RETURN_NAMES = ("CONDITIONING","label",)
+
+    FUNCTION = "run"
+    # INPUT_IS_LIST = True
+    CATEGORY = "♾️Mixlab/Prompt"
+
+    def run(self, conditioning, clip, gligen_textbox_model, grids, labels, index,max_size,random_shuffle,seed=0):
+        # print('grids',grids)
+        # conditioning=conditioning[0]
+        # clip=clip[0]
+        # gligen_textbox_model=gligen_textbox_model[0]
+        # index=index[0]
+        # max_size=max_size[0]
+        # random_shuffle=random_shuffle[0]
+
+        texts=labels
+        
+        if index>-1:
+            texts=[labels[index]]
+            grids=[grids[index]]
+
+        if random_shuffle=='on':
+            sss=[[texts[i],grids[i]] for i in range(len(texts))]
+            random.shuffle(sss)
+            texts=[s[0] for s in sss]
+            grids=[s[1] for s in sss]
+
+        if len(texts) > max_size:
+            texts = texts[:max_size]
+
+        c = []
+        
+        for t in conditioning:
+            n = [t[0], t[1].copy()]
+
+
+            # 多个
+            position_params=[]
+            for i in range(len(texts)):
+                text=texts[i]
+                grid=grids[i]
+                x,y,width,height=grid
+                # print(text)
+                cond, cond_pooled = clip.encode_from_tokens(clip.tokenize(text), return_pooled=True)
+                position_params =position_params+ [(cond_pooled, height // 8, width // 8, y // 8, x // 8)]
+
+            # 前一个
+            prev = []
+            if "gligen" in n[1]:
+                prev = n[1]['gligen'][2]
+
+            n[1]['gligen'] = ("position", gligen_textbox_model, prev + position_params)
+            # print('gligen',n)
+            c.append(n)
+        
+        #  下面这个写法有bug
+        # for i in range(len(texts)):
+        #     text=texts[i]
+        #     grid=grids[i]
+        #     x,y,width,height=grid
+
+        #     cond, cond_pooled = clip.encode_from_tokens(clip.tokenize(text), return_pooled=True)
+        #     for t in conditioning:
+        #         n = [t[0], t[1].copy()]
+        #         position_params = [(cond_pooled, height // 8, width // 8, y // 8, x // 8)]
+        #         prev = []
+        #         if "gligen" in n[1]:
+        #             prev = n[1]['gligen'][2]
+
+        #         n[1]['gligen'] = ("position", gligen_textbox_model, prev + position_params)
+        #         c.append(n)
+
+
+        return (c,texts, )
+    
+
+class JoinWithDelimiter:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+                    "text_list": (any_type,),
+                    "delimiter":(["newline","comma","backslash","space"],),
+                             },
+                }
+    
+    RETURN_TYPES = ("STRING",) 
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Text"
+
+    INPUT_IS_LIST = True # 当true的时候，输入时list，当false的时候，如果输入是list，则会自动包一层for循环调用
+    OUTPUT_IS_LIST = (False,)
+
+    def run(self,text_list,delimiter):
+        delimiter=delimiter[0]
+        if delimiter =='newline':
+            delimiter='\n'
+        elif delimiter=='comma':
+            delimiter=','
+        elif delimiter=='backslash':
+            delimiter='\\'
+        elif delimiter=='space':
+            delimiter=' '
+        t=''
+        if isinstance(text_list, list):
+            t=join_with_(text_list,delimiter)
+        return (t,)
@@ -0,0 +1,708 @@
+import os,sys
+import folder_paths
+
+from PIL import Image
+import importlib.util
+
+import comfy.utils
+import numpy as np
+import torch
+
+from huggingface_hub import hf_hub_download
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.transforms.functional import normalize
+# BRIA-RMBG-1.4 / briarmbg.py
+class REBNCONV(nn.Module):
+    def __init__(self,in_ch=3,out_ch=3,dirate=1,stride=1):
+        super(REBNCONV,self).__init__()
+
+        self.conv_s1 = nn.Conv2d(in_ch,out_ch,3,padding=1*dirate,dilation=1*dirate,stride=stride)
+        self.bn_s1 = nn.BatchNorm2d(out_ch)
+        self.relu_s1 = nn.ReLU(inplace=True)
+
+    def forward(self,x):
+
+        hx = x
+        xout = self.relu_s1(self.bn_s1(self.conv_s1(hx)))
+
+        return xout
+
+## upsample tensor 'src' to have the same spatial size with tensor 'tar'
+def _upsample_like(src,tar):
+
+    src = F.interpolate(src,size=tar.shape[2:],mode='bilinear')
+
+    return src
+
+
+### RSU-7 ###
+class RSU7(nn.Module):
+
+    def __init__(self, in_ch=3, mid_ch=12, out_ch=3, img_size=512):
+        super(RSU7,self).__init__()
+
+        self.in_ch = in_ch
+        self.mid_ch = mid_ch
+        self.out_ch = out_ch
+
+        self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1) ## 1 -> 1/2
+
+        self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1)
+        self.pool1 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=1)
+        self.pool2 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=1)
+        self.pool3 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=1)
+        self.pool4 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.rebnconv5 = REBNCONV(mid_ch,mid_ch,dirate=1)
+        self.pool5 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.rebnconv6 = REBNCONV(mid_ch,mid_ch,dirate=1)
+
+        self.rebnconv7 = REBNCONV(mid_ch,mid_ch,dirate=2)
+
+        self.rebnconv6d = REBNCONV(mid_ch*2,mid_ch,dirate=1)
+        self.rebnconv5d = REBNCONV(mid_ch*2,mid_ch,dirate=1)
+        self.rebnconv4d = REBNCONV(mid_ch*2,mid_ch,dirate=1)
+        self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=1)
+        self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=1)
+        self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1)
+
+    def forward(self,x):
+        b, c, h, w = x.shape
+
+        hx = x
+        hxin = self.rebnconvin(hx)
+
+        hx1 = self.rebnconv1(hxin)
+        hx = self.pool1(hx1)
+
+        hx2 = self.rebnconv2(hx)
+        hx = self.pool2(hx2)
+
+        hx3 = self.rebnconv3(hx)
+        hx = self.pool3(hx3)
+
+        hx4 = self.rebnconv4(hx)
+        hx = self.pool4(hx4)
+
+        hx5 = self.rebnconv5(hx)
+        hx = self.pool5(hx5)
+
+        hx6 = self.rebnconv6(hx)
+
+        hx7 = self.rebnconv7(hx6)
+
+        hx6d =  self.rebnconv6d(torch.cat((hx7,hx6),1))
+        hx6dup = _upsample_like(hx6d,hx5)
+
+        hx5d =  self.rebnconv5d(torch.cat((hx6dup,hx5),1))
+        hx5dup = _upsample_like(hx5d,hx4)
+
+        hx4d = self.rebnconv4d(torch.cat((hx5dup,hx4),1))
+        hx4dup = _upsample_like(hx4d,hx3)
+
+        hx3d = self.rebnconv3d(torch.cat((hx4dup,hx3),1))
+        hx3dup = _upsample_like(hx3d,hx2)
+
+        hx2d = self.rebnconv2d(torch.cat((hx3dup,hx2),1))
+        hx2dup = _upsample_like(hx2d,hx1)
+
+        hx1d = self.rebnconv1d(torch.cat((hx2dup,hx1),1))
+
+        return hx1d + hxin
+
+
+### RSU-6 ###
+class RSU6(nn.Module):
+
+    def __init__(self, in_ch=3, mid_ch=12, out_ch=3):
+        super(RSU6,self).__init__()
+
+        self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1)
+
+        self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1)
+        self.pool1 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=1)
+        self.pool2 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=1)
+        self.pool3 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=1)
+        self.pool4 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.rebnconv5 = REBNCONV(mid_ch,mid_ch,dirate=1)
+
+        self.rebnconv6 = REBNCONV(mid_ch,mid_ch,dirate=2)
+
+        self.rebnconv5d = REBNCONV(mid_ch*2,mid_ch,dirate=1)
+        self.rebnconv4d = REBNCONV(mid_ch*2,mid_ch,dirate=1)
+        self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=1)
+        self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=1)
+        self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1)
+
+    def forward(self,x):
+
+        hx = x
+
+        hxin = self.rebnconvin(hx)
+
+        hx1 = self.rebnconv1(hxin)
+        hx = self.pool1(hx1)
+
+        hx2 = self.rebnconv2(hx)
+        hx = self.pool2(hx2)
+
+        hx3 = self.rebnconv3(hx)
+        hx = self.pool3(hx3)
+
+        hx4 = self.rebnconv4(hx)
+        hx = self.pool4(hx4)
+
+        hx5 = self.rebnconv5(hx)
+
+        hx6 = self.rebnconv6(hx5)
+
+
+        hx5d =  self.rebnconv5d(torch.cat((hx6,hx5),1))
+        hx5dup = _upsample_like(hx5d,hx4)
+
+        hx4d = self.rebnconv4d(torch.cat((hx5dup,hx4),1))
+        hx4dup = _upsample_like(hx4d,hx3)
+
+        hx3d = self.rebnconv3d(torch.cat((hx4dup,hx3),1))
+        hx3dup = _upsample_like(hx3d,hx2)
+
+        hx2d = self.rebnconv2d(torch.cat((hx3dup,hx2),1))
+        hx2dup = _upsample_like(hx2d,hx1)
+
+        hx1d = self.rebnconv1d(torch.cat((hx2dup,hx1),1))
+
+        return hx1d + hxin
+
+### RSU-5 ###
+class RSU5(nn.Module):
+
+    def __init__(self, in_ch=3, mid_ch=12, out_ch=3):
+        super(RSU5,self).__init__()
+
+        self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1)
+
+        self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1)
+        self.pool1 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=1)
+        self.pool2 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=1)
+        self.pool3 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=1)
+
+        self.rebnconv5 = REBNCONV(mid_ch,mid_ch,dirate=2)
+
+        self.rebnconv4d = REBNCONV(mid_ch*2,mid_ch,dirate=1)
+        self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=1)
+        self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=1)
+        self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1)
+
+    def forward(self,x):
+
+        hx = x
+
+        hxin = self.rebnconvin(hx)
+
+        hx1 = self.rebnconv1(hxin)
+        hx = self.pool1(hx1)
+
+        hx2 = self.rebnconv2(hx)
+        hx = self.pool2(hx2)
+
+        hx3 = self.rebnconv3(hx)
+        hx = self.pool3(hx3)
+
+        hx4 = self.rebnconv4(hx)
+
+        hx5 = self.rebnconv5(hx4)
+
+        hx4d = self.rebnconv4d(torch.cat((hx5,hx4),1))
+        hx4dup = _upsample_like(hx4d,hx3)
+
+        hx3d = self.rebnconv3d(torch.cat((hx4dup,hx3),1))
+        hx3dup = _upsample_like(hx3d,hx2)
+
+        hx2d = self.rebnconv2d(torch.cat((hx3dup,hx2),1))
+        hx2dup = _upsample_like(hx2d,hx1)
+
+        hx1d = self.rebnconv1d(torch.cat((hx2dup,hx1),1))
+
+        return hx1d + hxin
+
+### RSU-4 ###
+class RSU4(nn.Module):
+
+    def __init__(self, in_ch=3, mid_ch=12, out_ch=3):
+        super(RSU4,self).__init__()
+
+        self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1)
+
+        self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1)
+        self.pool1 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=1)
+        self.pool2 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=1)
+
+        self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=2)
+
+        self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=1)
+        self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=1)
+        self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1)
+
+    def forward(self,x):
+
+        hx = x
+
+        hxin = self.rebnconvin(hx)
+
+        hx1 = self.rebnconv1(hxin)
+        hx = self.pool1(hx1)
+
+        hx2 = self.rebnconv2(hx)
+        hx = self.pool2(hx2)
+
+        hx3 = self.rebnconv3(hx)
+
+        hx4 = self.rebnconv4(hx3)
+
+        hx3d = self.rebnconv3d(torch.cat((hx4,hx3),1))
+        hx3dup = _upsample_like(hx3d,hx2)
+
+        hx2d = self.rebnconv2d(torch.cat((hx3dup,hx2),1))
+        hx2dup = _upsample_like(hx2d,hx1)
+
+        hx1d = self.rebnconv1d(torch.cat((hx2dup,hx1),1))
+
+        return hx1d + hxin
+
+### RSU-4F ###
+class RSU4F(nn.Module):
+
+    def __init__(self, in_ch=3, mid_ch=12, out_ch=3):
+        super(RSU4F,self).__init__()
+
+        self.rebnconvin = REBNCONV(in_ch,out_ch,dirate=1)
+
+        self.rebnconv1 = REBNCONV(out_ch,mid_ch,dirate=1)
+        self.rebnconv2 = REBNCONV(mid_ch,mid_ch,dirate=2)
+        self.rebnconv3 = REBNCONV(mid_ch,mid_ch,dirate=4)
+
+        self.rebnconv4 = REBNCONV(mid_ch,mid_ch,dirate=8)
+
+        self.rebnconv3d = REBNCONV(mid_ch*2,mid_ch,dirate=4)
+        self.rebnconv2d = REBNCONV(mid_ch*2,mid_ch,dirate=2)
+        self.rebnconv1d = REBNCONV(mid_ch*2,out_ch,dirate=1)
+
+    def forward(self,x):
+
+        hx = x
+
+        hxin = self.rebnconvin(hx)
+
+        hx1 = self.rebnconv1(hxin)
+        hx2 = self.rebnconv2(hx1)
+        hx3 = self.rebnconv3(hx2)
+
+        hx4 = self.rebnconv4(hx3)
+
+        hx3d = self.rebnconv3d(torch.cat((hx4,hx3),1))
+        hx2d = self.rebnconv2d(torch.cat((hx3d,hx2),1))
+        hx1d = self.rebnconv1d(torch.cat((hx2d,hx1),1))
+
+        return hx1d + hxin
+
+
+class myrebnconv(nn.Module):
+    def __init__(self, in_ch=3,
+                       out_ch=1,
+                       kernel_size=3,
+                       stride=1,
+                       padding=1,
+                       dilation=1,
+                       groups=1):
+        super(myrebnconv,self).__init__()
+
+        self.conv = nn.Conv2d(in_ch,
+                              out_ch,
+                              kernel_size=kernel_size,
+                              stride=stride,
+                              padding=padding,
+                              dilation=dilation,
+                              groups=groups)
+        self.bn = nn.BatchNorm2d(out_ch)
+        self.rl = nn.ReLU(inplace=True)
+
+    def forward(self,x):
+        return self.rl(self.bn(self.conv(x)))
+
+
+class BriaRMBG(nn.Module):
+
+    def __init__(self,in_ch=3,out_ch=1):
+        super(BriaRMBG,self).__init__()
+
+        self.conv_in = nn.Conv2d(in_ch,64,3,stride=2,padding=1)
+        self.pool_in = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.stage1 = RSU7(64,32,64)
+        self.pool12 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.stage2 = RSU6(64,32,128)
+        self.pool23 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.stage3 = RSU5(128,64,256)
+        self.pool34 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.stage4 = RSU4(256,128,512)
+        self.pool45 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.stage5 = RSU4F(512,256,512)
+        self.pool56 = nn.MaxPool2d(2,stride=2,ceil_mode=True)
+
+        self.stage6 = RSU4F(512,256,512)
+
+        # decoder
+        self.stage5d = RSU4F(1024,256,512)
+        self.stage4d = RSU4(1024,128,256)
+        self.stage3d = RSU5(512,64,128)
+        self.stage2d = RSU6(256,32,64)
+        self.stage1d = RSU7(128,16,64)
+
+        self.side1 = nn.Conv2d(64,out_ch,3,padding=1)
+        self.side2 = nn.Conv2d(64,out_ch,3,padding=1)
+        self.side3 = nn.Conv2d(128,out_ch,3,padding=1)
+        self.side4 = nn.Conv2d(256,out_ch,3,padding=1)
+        self.side5 = nn.Conv2d(512,out_ch,3,padding=1)
+        self.side6 = nn.Conv2d(512,out_ch,3,padding=1)
+
+        # self.outconv = nn.Conv2d(6*out_ch,out_ch,1)
+
+    def forward(self,x):
+
+        hx = x
+
+        hxin = self.conv_in(hx)
+        #hx = self.pool_in(hxin)
+
+        #stage 1
+        hx1 = self.stage1(hxin)
+        hx = self.pool12(hx1)
+
+        #stage 2
+        hx2 = self.stage2(hx)
+        hx = self.pool23(hx2)
+
+        #stage 3
+        hx3 = self.stage3(hx)
+        hx = self.pool34(hx3)
+
+        #stage 4
+        hx4 = self.stage4(hx)
+        hx = self.pool45(hx4)
+
+        #stage 5
+        hx5 = self.stage5(hx)
+        hx = self.pool56(hx5)
+
+        #stage 6
+        hx6 = self.stage6(hx)
+        hx6up = _upsample_like(hx6,hx5)
+
+        #-------------------- decoder --------------------
+        hx5d = self.stage5d(torch.cat((hx6up,hx5),1))
+        hx5dup = _upsample_like(hx5d,hx4)
+
+        hx4d = self.stage4d(torch.cat((hx5dup,hx4),1))
+        hx4dup = _upsample_like(hx4d,hx3)
+
+        hx3d = self.stage3d(torch.cat((hx4dup,hx3),1))
+        hx3dup = _upsample_like(hx3d,hx2)
+
+        hx2d = self.stage2d(torch.cat((hx3dup,hx2),1))
+        hx2dup = _upsample_like(hx2d,hx1)
+
+        hx1d = self.stage1d(torch.cat((hx2dup,hx1),1))
+
+
+        #side output
+        d1 = self.side1(hx1d)
+        d1 = _upsample_like(d1,x)
+
+        d2 = self.side2(hx2d)
+        d2 = _upsample_like(d2,x)
+
+        d3 = self.side3(hx3d)
+        d3 = _upsample_like(d3,x)
+
+        d4 = self.side4(hx4d)
+        d4 = _upsample_like(d4,x)
+
+        d5 = self.side5(hx5d)
+        d5 = _upsample_like(d5,x)
+
+        d6 = self.side6(hx6)
+        d6 = _upsample_like(d6,x)
+
+        return [F.sigmoid(d1), F.sigmoid(d2), F.sigmoid(d3), F.sigmoid(d4), F.sigmoid(d5), F.sigmoid(d6)],[hx1d,hx2d,hx3d,hx4d,hx5d,hx6]
+    
+
+
+
+def get_U2NET_model_path():
+    try:
+        return folder_paths.get_folder_paths('rembg')[0]
+    except:
+        return os.path.join(folder_paths.models_dir, "rembg")
+    
+
+U2NET_HOME=get_U2NET_model_path()
+os.environ["U2NET_HOME"] = U2NET_HOME
+
+global _available
+_available=False
+
+
+def get_rembg_models(path):
+    """从目录中获取文件并提取文件名
+    Args:
+        path: 目录路径
+    Returns:
+        文件名列表
+    """
+    filenames = []
+    for root, _, files in os.walk(path):
+        for filename in files:
+        # 过滤隐藏文件
+            if not filename.startswith('.'):
+                name, ext = os.path.splitext(os.path.basename(filename))
+                filenames.append(name)
+    return filenames
+
+
+def is_installed(package):
+    try:
+        spec = importlib.util.find_spec(package)
+    except ModuleNotFoundError:
+        return False
+    return spec is not None
+
+
+try:
+    if is_installed('rembg')==False:
+        import subprocess
+
+        # 安装
+        print('#pip install rembg[gpu]')
+
+        result = subprocess.run([sys.executable, '-s', '-m', 'pip', 'install', 'rembg[gpu]'], capture_output=True, text=True)
+
+        #检查命令执行结果
+        if result.returncode == 0:
+            print("#install success")
+            from rembg import new_session, remove
+            _available=True
+        else:
+            print("#install error")
+        
+    else:
+        from rembg import new_session, remove
+        _available=True
+
+except:
+    _available=False
+
+
+def run_briarmbg(images=[]):
+    mroot=U2NET_HOME
+    m=os.path.join(mroot,'briarmbg.pth')
+    if os.path.exists(m)==False:
+        # 下载
+        m1=hf_hub_download("briaai/RMBG-1.4",
+                       local_dir=mroot,
+                       filename='model.pth',
+                       local_dir_use_symlinks=False,
+                       endpoint='https://hf-mirror.com')
+        os.rename(m1, m)
+    
+    net=BriaRMBG()
+    if torch.cuda.is_available():
+        net.load_state_dict(torch.load(m))
+        net=net.cuda()
+    else:
+        net.load_state_dict(torch.load(m,map_location="cpu"))
+    net.eval() 
+
+    masks=[]
+    rgba_images=[]
+    rgb_images=[]
+    for orig_image in images:
+
+        w,h = orig_im_size = orig_image.size
+
+        image = orig_image.convert('RGB')
+        model_input_size = (1024, 1024)
+        image = image.resize(model_input_size, Image.BILINEAR)
+
+        im_np = np.array(image)
+        im_tensor = torch.tensor(im_np, dtype=torch.float32).permute(2,0,1)
+        im_tensor = torch.unsqueeze(im_tensor,0)
+        im_tensor = torch.divide(im_tensor,255.0)
+        im_tensor = normalize(im_tensor,[0.5,0.5,0.5],[1.0,1.0,1.0])
+        if torch.cuda.is_available():
+            im_tensor=im_tensor.cuda()
+
+        result=net(im_tensor)
+        result = torch.squeeze(F.interpolate(result[0][0], size=(h,w), mode='bilinear') ,0)
+        ma = torch.max(result)
+        mi = torch.min(result)
+        result = (result-mi)/(ma-mi)
+        im_array = (result*255).cpu().data.numpy().astype(np.uint8)
+        mask = Image.fromarray(np.squeeze(im_array))
+        # mask.save('test.png')
+        # mask=tensor2pil(result)
+        mask=mask.convert('L') 
+
+        masks.append(mask)
+
+        # rgba图
+        image_rgba =orig_image.convert("RGBA")
+        image_rgba.putalpha(mask)
+        rgba_images.append(image_rgba)
+
+        #rgb
+        rgb_image = Image.new("RGB", image_rgba.size, (0, 0, 0))
+        rgb_image.paste(image_rgba, mask=image_rgba.split()[3])
+        rgb_images.append(rgb_image)
+    return (masks,rgba_images,rgb_images)   
+
+
+def run_rembg(model_name= "unet",images=[],callback=None):
+    # model_name = "unet" # "isnet-general-use"
+    # print('#run_rembg',model_name)
+    rembg_session = new_session(model_name)
+    masks=[]
+    rgba_images=[]
+    rgb_images=[]
+    # 进度条
+    pbar=callback 
+    for img in images:
+        # use the post_process_mask argument to post process the mask to get better results.
+        mask = remove(img, session=rembg_session,only_mask=True,post_process_mask=True)
+        # mask=mask.convert('L')
+        # masks.append(mask)
+        if model_name=="u2net_cloth_seg":
+            width, original_height = mask.size
+            num_slices = original_height // img.height
+            for i in range(num_slices):
+                top = i * img.height
+                bottom = (i + 1) * img.height
+                slice_image = mask.crop((0, top, width, bottom))
+                slice_mask=slice_image.convert('L')
+                masks.append(slice_mask)
+
+                # rgba图
+                image_rgba = img.convert("RGBA")
+                image_rgba.putalpha(slice_mask)
+                rgba_images.append(image_rgba)
+
+                #rgb
+                rgb_image = Image.new("RGB", image_rgba.size, (0, 0, 0))
+                rgb_image.paste(image_rgba, mask=image_rgba.split()[3])
+                rgb_images.append(rgb_image)
+
+        else:
+            mask=mask.convert('L')    
+            # mask.save(output_path)
+            masks.append(mask)
+
+            # rgba图
+            image_rgba = img.convert("RGBA")
+            image_rgba.putalpha(mask)
+            rgba_images.append(image_rgba)
+
+            #rgb
+            rgb_image = Image.new("RGB", image_rgba.size, (0, 0, 0))
+            rgb_image.paste(image_rgba, mask=image_rgba.split()[3])
+            rgb_images.append(rgb_image)
+
+        if pbar:
+            pbar.update(1)
+    return (masks,rgba_images,rgb_images)
+
+
+# Tensor to PIL
+def tensor2pil(image):
+    return Image.fromarray(np.clip(255. * image.cpu().numpy().squeeze(), 0, 255).astype(np.uint8))
+
+# Convert PIL to Tensor
+def pil2tensor(image):
+    return torch.from_numpy(np.array(image).astype(np.float32) / 255.0).unsqueeze(0)
+
+
+class RembgNode_:
+
+    global _available
+    available=_available
+    
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "image": ("IMAGE",),
+            "model_name": (get_rembg_models(U2NET_HOME),),
+            
+                             },
+                }
+    
+    RETURN_TYPES = ("MASK","IMAGE","RGBA",)
+    RETURN_NAMES = ("masks","images","RGBAs")
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Mask"
+    OUTPUT_NODE = True
+    INPUT_IS_LIST = True
+    OUTPUT_IS_LIST = (True,True,True,)
+  
+    def run(self,image,model_name):
+        # 兼容list输入和batch输入
+
+        model_name=model_name[0]
+
+        images=[]
+        
+        for ims in image:
+            for im in ims:
+                im=tensor2pil(im)
+                images.append(im)
+
+        if model_name=='briarmbg':
+            masks,rgba_images,rgb_images=run_briarmbg(images)
+        else:
+            masks,rgba_images,rgb_images=run_rembg(model_name,images, comfy.utils.ProgressBar(len(images) ))
+
+        masks=[pil2tensor(m) for m in masks]
+
+        rgba_images=[pil2tensor(m) for m in rgba_images]
+
+        rgb_images=[pil2tensor(m) for m in rgb_images]
+
+        return (masks,rgb_images,rgba_images,)
@@ -79,35 +79,37 @@ class ScreenShareNode:
    def INPUT_TYPES(s):
        return { "required":{
            "image_base64": ("CHEESE",),
+            "refresh_rate": ("INT", {"default": 500, "min": 0,"step": 50, "max": 0xffffffffffffffff}),
        },
          "optional":{
              "prompt": ("PROMPT",),
              "slide": ("SLIDE",),
              "seed": ("SEED",),
+              
            #   "seed": ("INT", {"default": 1, "min": 0, "max": 0xffffffffffffffff}),
                } }
    
    RETURN_TYPES = ('IMAGE','STRING','FLOAT',"INT")
-    RETURN_NAMES = ("IMAGE","PROMPT","FLOAT","INT")
+    RETURN_NAMES = ("current frame (image)","prompt","denoise (float)","seed (int)")
    FUNCTION = "run"

-    CATEGORY = "♾️Mixlab/image"
+    CATEGORY = "♾️Mixlab/Screen"

    # INPUT_IS_LIST = True
    OUTPUT_IS_LIST = (False,False,False,False)
  
    # 运行的函数
-    def run(self,image_base64,prompt,slide,seed):
+    def run(self,image_base64,refresh_rate ,prompt,slide,seed):
        im,mask=base64_save(image_base64)
        # print('##########prompt',prompt)
-        return (im,prompt,slide,seed)
-    
+        return {"ui":{"refresh_rate": [refresh_rate]},"result": (im,prompt,slide,seed,)}
+

 class FloatingVideo:
    @classmethod
    def INPUT_TYPES(s):
        return { "required":{
-            "images": ("IMAGE",)
+            "image": ("IMAGE",)
        }, }
    
    # RETURN_TYPES = ('IMAGE','MASK')
@@ -116,22 +118,22 @@ class FloatingVideo:
    OUTPUT_NODE = True
    FUNCTION = "run"

-    CATEGORY = "♾️Mixlab/image"
+    CATEGORY = "♾️Mixlab/Screen"

    # INPUT_IS_LIST = True
    # OUTPUT_IS_LIST = (False,False,)
  
    # 运行的函数
-    def run(self,images):
+    def run(self,image):
       
        results = list()
    
-        for image in images:
-            image=tensor2pil(image)
+        for im in image:
+            im=tensor2pil(im)
            # image_base64 = base64.b64encode(image.tobytes())

            buffered = BytesIO()
-            image.save(buffered, format="JPEG")
+            im.save(buffered, format="JPEG")
            image_base64 = base64.b64encode(buffered.getvalue()).decode("utf-8")

            results.append(image_base64)
@@ -0,0 +1,503 @@
+import comfy
+import torch
+
+from dataclasses import dataclass
+import torch.nn as nn
+from comfy.model_patcher import ModelPatcher
+import comfy.ops
+from typing import Union
+import comfy.sample
+import latent_preview
+import comfy.utils
+
+T = torch.Tensor
+
+
+from .VisualStylePrompting.attention_functions import VisualStyleProcessor
+
+class ApplyVisualStylePrompting:
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "reference_image": ("IMAGE",),
+                "reference_image_text": ("STRING", {"multiline": True}),
+                "model": ("MODEL",),
+                "clip": ("CLIP", ),
+                "vae": ("VAE", ),
+                "positive": ("CONDITIONING",),
+                "negative": ("CONDITIONING", ),
+                "enabled": ("BOOLEAN", {"default": True}),
+                "denoise": ("FLOAT", {"default": 1., "min": 0., "max": 1., "step": 1e-2}),
+                "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096,"step":2})
+            }
+        }
+
+    RETURN_TYPES = ("MODEL", "CONDITIONING","CONDITIONING", "LATENT")
+    RETURN_NAMES = ("model", "positive", "negative", "latents")
+    
+    CATEGORY = "♾️Mixlab/Style"
+
+    FUNCTION = "run"
+
+    def run(
+        self, 
+        reference_image, 
+        reference_image_text,
+        model: comfy.model_patcher.ModelPatcher, 
+        clip,
+        vae,
+        positive,
+        negative, 
+        enabled,
+        denoise,
+        batch_size=1
+    ):
+        
+        tokens = clip.tokenize(reference_image_text)
+        cond, pooled = clip.encode_from_tokens(tokens, return_pooled=True)
+        reference_image_prompt=[[cond, {"pooled_output": pooled}]]
+   
+        reference_image = reference_image.repeat(((batch_size+1)//2, 1,1,1))
+
+        self.model = model
+        reference_latent = vae.encode(reference_image[:,:,:,:3])
+        
+        for n, m in model.model.diffusion_model.named_modules():
+            if m.__class__.__name__  == "CrossAttention":
+                processor = VisualStyleProcessor(m, enabled=enabled)
+                setattr(m, 'forward', processor.visual_style_forward)
+
+        conditioning_prompt = reference_image_prompt + positive
+        negative_prompt = negative * 2 
+
+        latents = torch.zeros_like(reference_latent) 
+        latents = torch.cat([latents] * 2)
+
+        if denoise < 1.0:
+            latents[::1] = reference_latent[:1]
+        else:
+            latents[::2] = reference_latent
+
+        denoise_mask = torch.ones_like(latents)[:, :1, ...] * denoise
+
+        denoise_mask[0] = 0.
+
+        return (model, conditioning_prompt, negative_prompt, {"samples": latents, "noise_mask": denoise_mask})    
+
+
+
+
+
+def exists(val):
+    return val is not None
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d
+
+
+class StyleAlignedArgs:
+    def __init__(self, share_attn: str) -> None:
+        self.adain_keys = "k" in share_attn
+        self.adain_values = "v" in share_attn
+        self.adain_queries = "q" in share_attn
+
+    share_attention: bool = True
+    adain_queries: bool = True
+    adain_keys: bool = True
+    adain_values: bool = True
+
+
+def expand_first(
+    feat: T,
+    scale=1.0,
+) -> T:
+    """
+    Expand the first element so it has the same shape as the rest of the batch.
+    """
+    b = feat.shape[0]
+    feat_style = torch.stack((feat[0], feat[b // 2])).unsqueeze(1)
+    if scale == 1:
+        feat_style = feat_style.expand(2, b // 2, *feat.shape[1:])
+    else:
+        feat_style = feat_style.repeat(1, b // 2, 1, 1, 1)
+        feat_style = torch.cat([feat_style[:, :1], scale * feat_style[:, 1:]], dim=1)
+    return feat_style.reshape(*feat.shape)
+
+
+def concat_first(feat: T, dim=2, scale=1.0) -> T:
+    """
+    concat the the feature and the style feature expanded above
+    """
+    feat_style = expand_first(feat, scale=scale)
+    return torch.cat((feat, feat_style), dim=dim)
+
+
+def calc_mean_std(feat, eps: float = 1e-5) -> "tuple[T, T]":
+    feat_std = (feat.var(dim=-2, keepdims=True) + eps).sqrt()
+    feat_mean = feat.mean(dim=-2, keepdims=True)
+    return feat_mean, feat_std
+
+def adain(feat: T) -> T:
+    feat_mean, feat_std = calc_mean_std(feat)
+    feat_style_mean = expand_first(feat_mean)
+    feat_style_std = expand_first(feat_std)
+    feat = (feat - feat_mean) / feat_std
+    feat = feat * feat_style_std + feat_style_mean
+    return feat
+
+class SharedAttentionProcessor:
+    def __init__(self, args: StyleAlignedArgs, scale: float):
+        self.args = args
+        self.scale = scale
+
+    def __call__(self, q, k, v, extra_options):
+        if self.args.adain_queries:
+            q = adain(q)
+        if self.args.adain_keys:
+            k = adain(k)
+        if self.args.adain_values:
+            v = adain(v)
+        if self.args.share_attention:
+            k = concat_first(k, -2, scale=self.scale)
+            v = concat_first(v, -2)
+
+        return q, k, v
+
+
+def get_norm_layers(
+    layer: nn.Module,
+    norm_layers_: "dict[str, list[Union[nn.GroupNorm, nn.LayerNorm]]]",
+    share_layer_norm: bool,
+    share_group_norm: bool,
+):
+    if isinstance(layer, nn.LayerNorm) and share_layer_norm:
+        norm_layers_["layer"].append(layer)
+    if isinstance(layer, nn.GroupNorm) and share_group_norm:
+        norm_layers_["group"].append(layer)
+    else:
+        for child_layer in layer.children():
+            get_norm_layers(
+                child_layer, norm_layers_, share_layer_norm, share_group_norm
+            )
+
+
+def register_norm_forward(
+    norm_layer: Union[nn.GroupNorm, nn.LayerNorm],
+) -> Union[nn.GroupNorm, nn.LayerNorm]:
+    if not hasattr(norm_layer, "orig_forward"):
+        setattr(norm_layer, "orig_forward", norm_layer.forward)
+    orig_forward = norm_layer.orig_forward
+
+    def forward_(hidden_states: T) -> T:
+        n = hidden_states.shape[-2]
+        hidden_states = concat_first(hidden_states, dim=-2)
+        hidden_states = orig_forward(hidden_states)  # type: ignore
+        return hidden_states[..., :n, :]
+
+    norm_layer.forward = forward_  # type: ignore
+    return norm_layer
+
+
+def register_shared_norm(
+    model: ModelPatcher,
+    share_group_norm: bool = True,
+    share_layer_norm: bool = True,
+):
+    norm_layers = {"group": [], "layer": []}
+    get_norm_layers(model.model, norm_layers, share_layer_norm, share_group_norm)
+    print(
+        f"Patching {len(norm_layers['group'])} group norms, {len(norm_layers['layer'])} layer norms."
+    )
+    return [register_norm_forward(layer) for layer in norm_layers["group"]] + [
+        register_norm_forward(layer) for layer in norm_layers["layer"]
+    ]
+
+
+SHARE_NORM_OPTIONS = ["both", "group", "layer", "disabled"]
+SHARE_ATTN_OPTIONS = ["q+k", "q+k+v", "disabled"]
+
+class StyleAlignedSampleReferenceLatents:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    { 
+                    "reference_image": ("IMAGE",),
+                    "positive": ("CONDITIONING",),
+                    "negative": ("CONDITIONING", ),
+                    "model": ("MODEL",),
+                    "vae": ("VAE", ),
+                    "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
+                    "steps": ("INT", {"default": 20, "min": 1, "max": 10000}),
+                    "cfg": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}),
+
+                    "scheduler": (comfy.samplers.KSampler.SCHEDULERS.reverse(), ), 
+                    "denoise": ("FLOAT", {"default": 1, "min": 0.0, "max": 1.0, "step": 0.01}),
+                    
+                     }
+                }
+
+    RETURN_TYPES = ("STEP_LATENTS","LATENT")
+    RETURN_NAMES = ("ref_latents", "noised_output")
+
+    FUNCTION = "run"
+
+    # CATEGORY = "style_aligned"
+    CATEGORY = "♾️Mixlab/Style"
+
+    def run(self, reference_image, positive, negative, model, vae, seed, steps, cfg,scheduler,denoise):
+
+        # TODO noise_mask?
+        def vae_encode_crop_pixels(pixels):
+            x = (pixels.shape[1] // 8) * 8
+            y = (pixels.shape[2] // 8) * 8
+            if pixels.shape[1] != x or pixels.shape[2] != y:
+                x_offset = (pixels.shape[1] % 8) // 2
+                y_offset = (pixels.shape[2] % 8) // 2
+                pixels = pixels[:, x_offset:x + x_offset, y_offset:y + y_offset, :]
+            return pixels
+
+        pixels=vae_encode_crop_pixels(reference_image)
+        t = vae.encode(pixels[:,:,:,:3])
+        latent_image =  {"samples":t}
+
+        noise_seed=seed
+
+        sampler_name="ddim"
+        
+        sampler = comfy.samplers.sampler_object(sampler_name)
+
+        total_steps = steps
+        if denoise < 1.0:
+            total_steps = int(steps/denoise)
+
+        comfy.model_management.load_models_gpu([model])
+        sigmas = comfy.samplers.calculate_sigmas_scheduler(model.model, scheduler, total_steps).cpu()
+        sigmas = sigmas[-(steps + 1):]
+
+        sigmas = sigmas.flip(0)
+        if sigmas[0] == 0:
+            sigmas[0] = 0.0001
+
+        latent = latent_image
+        latent_image = latent["samples"]
+        noise = torch.zeros(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, device="cpu")
+
+    
+        noise_mask = None
+        if "noise_mask" in latent:
+            noise_mask = latent["noise_mask"]
+
+        ref_latents = []
+        def callback(step: int, x0: T, x: T, steps: int):
+            ref_latents.insert(0, x[0])
+        
+        disable_pbar = not comfy.utils.PROGRESS_BAR_ENABLED
+        samples = comfy.sample.sample_custom(model, noise, cfg, sampler, sigmas, positive, negative, latent_image, noise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=noise_seed)
+
+        out = latent.copy()
+        out["samples"] = samples
+        out_noised = out
+
+        ref_latents = torch.stack(ref_latents)
+
+        return (ref_latents, out_noised)
+
+class StyleAlignedReferenceSampler:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+
+                "ref_latents": ("STEP_LATENTS",),
+                "reference_image_text": ("STRING", {"multiline": True}),
+                "model": ("MODEL",),
+                "clip": ("CLIP", ),
+
+                "positive": ("CONDITIONING",),
+                "negative": ("CONDITIONING",),
+
+                "share_norm": (SHARE_NORM_OPTIONS,),
+                "share_attn": (SHARE_ATTN_OPTIONS,),
+                "scale": ("FLOAT", {"default": 1, "min": 0, "max": 2.0, "step": 0.01}),
+                "batch_size": ("INT", {"default": 2, "min": 1, "max": 8, "step": 1}),
+                "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
+                "steps": ("INT", {"default": 20, "min": 1, "max": 10000}),
+                "cfg": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}),
+                "scheduler": (comfy.samplers.KSampler.SCHEDULERS, ), 
+                "denoise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
+            },
+        }
+
+    RETURN_TYPES = ("LATENT", "LATENT")
+    RETURN_NAMES = ("output", "denoised_output")
+    FUNCTION = "patch"
+    # CATEGORY = "style_aligned"
+    CATEGORY = "♾️Mixlab/Style"
+    def patch(
+        self,
+        ref_latents,
+        reference_image_text,
+        model,
+        clip,
+        positive,
+        negative,
+        share_norm,
+        share_attn,
+        scale,
+        batch_size,
+        seed,steps,cfg,scheduler,denoise
+
+    ) -> "tuple[dict, dict]":
+        
+        m = model.clone()
+
+        # ref_latents = vae.encode(reference_image[:,:,:,:3])
+
+        tokens = clip.tokenize(reference_image_text)
+        cond, pooled = clip.encode_from_tokens(tokens, return_pooled=True)
+        ref_positive=[[cond, {"pooled_output": pooled}]]
+
+        noise_seed=seed
+
+
+        total_steps = steps
+        if denoise < 1.0:
+            total_steps = int(steps/denoise)
+
+        # comfy.model_management.load_models_gpu([model])
+        sigmas = comfy.samplers.calculate_sigmas_scheduler(model.model, scheduler, total_steps).cpu()
+        sigmas = sigmas[-(steps + 1):]
+
+        sampler_name="ddim"
+        
+        sampler = comfy.samplers.sampler_object(sampler_name)
+
+        args = StyleAlignedArgs(share_attn)
+
+        # Concat batch with style latent
+        style_latent_tensor = ref_latents[0].unsqueeze(0)
+        height, width = style_latent_tensor.shape[-2:]
+        latent_t = torch.zeros(
+            [batch_size, 4, height, width], device=ref_latents.device
+        )
+        latent = {"samples": latent_t}
+        noise = comfy.sample.prepare_noise(latent_t, noise_seed)
+
+        latent_t = torch.cat((style_latent_tensor, latent_t), dim=0)
+        ref_noise = torch.zeros_like(noise[0]).unsqueeze(0)
+        noise = torch.cat((ref_noise, noise), dim=0)
+
+        x0_output = {}
+        preview_callback = latent_preview.prepare_callback(m, sigmas.shape[-1] - 1, x0_output)
+
+        # Replace first latent with the corresponding reference latent after each step
+        def callback(step: int, x0: T, x: T, steps: int):
+            preview_callback(step, x0, x, steps)
+            if (step + 1 < steps):
+                # 当ref_latents的step不够时
+                if step+1>len(ref_latents)-1:
+                    step=len(ref_latents)-2
+                
+                x[0] = ref_latents[step+1]
+                x0[0] = ref_latents[step+1]
+
+        # Register shared norms
+        share_group_norm = share_norm in ["group", "both"]
+        share_layer_norm = share_norm in ["layer", "both"]
+        register_shared_norm(m, share_group_norm, share_layer_norm)
+
+        # Patch cross attn
+        m.set_model_attn1_patch(SharedAttentionProcessor(args, scale))
+
+        # Add reference conditioning to batch 
+        batched_condition = []
+        for i,condition in enumerate(positive):
+            additional = condition[1].copy()
+            batch_with_reference = torch.cat([ref_positive[i][0], condition[0].repeat([batch_size] + [1] * len(condition[0].shape[1:]))], dim=0)
+            if 'pooled_output' in additional and 'pooled_output' in ref_positive[i][1]:
+                # combine pooled output
+                pooled_output = torch.cat([ref_positive[i][1]['pooled_output'], additional['pooled_output'].repeat([batch_size] 
+                    + [1] * len(additional['pooled_output'].shape[1:]))], dim=0)
+                additional['pooled_output'] = pooled_output
+            if 'control' in additional:
+                if 'control' in ref_positive[i][1]:
+                    # combine control conditioning
+                    control_hint = torch.cat([ref_positive[i][1]['control'].cond_hint_original, additional['control'].cond_hint_original.repeat([batch_size] 
+                        + [1] * len(additional['control'].cond_hint_original.shape[1:]))], dim=0)
+                    cloned_controlnet = additional['control'].copy()
+                    cloned_controlnet.set_cond_hint(control_hint, strength=additional['control'].strength, timestep_percent_range=additional['control'].timestep_percent_range)
+                    additional['control'] = cloned_controlnet
+                else:
+                    # add zeros for first in batch
+                    control_hint = torch.cat([torch.zeros_like(additional['control'].cond_hint_original), additional['control'].cond_hint_original.repeat([batch_size] 
+                        + [1] * len(additional['control'].cond_hint_original.shape[1:]))], dim=0)
+                    cloned_controlnet = additional['control'].copy()
+                    cloned_controlnet.set_cond_hint(control_hint, strength=additional['control'].strength, timestep_percent_range=additional['control'].timestep_percent_range)
+                    additional['control'] = cloned_controlnet
+            batched_condition.append([batch_with_reference, additional])
+
+        disable_pbar = not comfy.utils.PROGRESS_BAR_ENABLED
+        samples = comfy.sample.sample_custom(
+            m,
+            noise,
+            cfg,
+            sampler,
+            sigmas,
+            batched_condition,
+            negative,
+            latent_t,
+            callback=callback,
+            disable_pbar=disable_pbar,
+            seed=noise_seed,
+        )
+
+        # remove reference image
+        samples = samples[1:]
+
+        out = latent.copy()
+        out["samples"] = samples
+        if "x0" in x0_output:
+            out_denoised = latent.copy()
+            x0 = x0_output["x0"][1:]
+            out_denoised["samples"] = m.model.process_latent_out(x0.cpu())
+        else:
+            out_denoised = out
+        return (out, out_denoised)
+
+
+class StyleAlignedBatchAlign:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "model": ("MODEL",),
+                "share_norm": (SHARE_NORM_OPTIONS,),
+                "share_attn": (SHARE_ATTN_OPTIONS,),
+                "scale": ("FLOAT", {"default": 1, "min": 0, "max": 1.0, "step": 0.1}),
+            }
+        }
+
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+    # CATEGORY = "style_aligned"
+    CATEGORY = "♾️Mixlab/Style"
+    def patch(
+        self,
+        model: ModelPatcher,
+        share_norm: str,
+        share_attn: str,
+        scale: float,
+    ):
+        m = model.clone()
+        share_group_norm = share_norm in ["group", "both"]
+        share_layer_norm = share_norm in ["layer", "both"]
+        register_shared_norm(model, share_group_norm, share_layer_norm)
+        args = StyleAlignedArgs(share_attn)
+        m.set_model_attn1_patch(SharedAttentionProcessor(args, scale))
+        return (m,)
+
+
@@ -0,0 +1,439 @@
+from transformers import pipeline, set_seed,AutoTokenizer, AutoModelForSeq2SeqLM
+import random
+import re
+
+import os,sys
+import folder_paths
+
+# from PIL import Image
+import importlib.util
+
+import comfy.utils
+# import numpy as np
+import torch
+import random
+from lark import Lark, Transformer, v_args
+
+
+global _available
+_available=True
+
+def get_text_generator_path():
+    try:
+        return folder_paths.get_folder_paths('prompt_generator')[0]
+    except:
+        return os.path.join(folder_paths.models_dir, "prompt_generator")
+
+prompt_generator=get_text_generator_path()
+
+text_generator_model_path=os.path.join(prompt_generator, "text2image-prompt-generator")
+if not os.path.exists(text_generator_model_path):
+    print(f"## text_generator_model not found: {text_generator_model_path}, pls download from https://huggingface.co/succinctly/text2image-prompt-generator/tree/main")
+    text_generator_model_path='succinctly/text2image-prompt-generator'
+
+zh_en_model_path=os.path.join(prompt_generator, "opus-mt-zh-en")
+if not os.path.exists(zh_en_model_path):
+    print(f"## zh_en_model not found: {zh_en_model_path}, pls download from https://huggingface.co/Helsinki-NLP/opus-mt-zh-en/tree/main")
+    zh_en_model_path='Helsinki-NLP/opus-mt-zh-en'
+
+
+def is_installed(package):
+    try:
+        spec = importlib.util.find_spec(package)
+    except ModuleNotFoundError:
+        return False
+    return spec is not None
+
+
+try:
+    if is_installed('sentencepiece')==False:
+        import subprocess
+
+        # 安装
+        print('#pip install sentencepiece')
+
+        result = subprocess.run([sys.executable, '-s', '-m', 'pip', 'install', 'sentencepiece'], capture_output=True, text=True)
+
+        #检查命令执行结果
+        if result.returncode == 0 and is_installed('sentencepiece'):
+            print("#install success")
+            _available=True
+        else:
+            print("#install error")
+            _available=False
+        
+    else:
+        _available=True
+
+except:
+    _available=False
+
+
+
+def translate(text):
+    global text_pipe,zh_en_model,zh_en_tokenizer
+    
+    if zh_en_model==None:
+        zh_en_model = AutoModelForSeq2SeqLM.from_pretrained(zh_en_model_path).eval()
+        zh_en_tokenizer = AutoTokenizer.from_pretrained(zh_en_model_path,padding=True, truncation=True)
+    
+    zh_en_model.to("cuda" if torch.cuda.is_available() else "cpu")
+    with torch.no_grad():
+        encoded = zh_en_tokenizer([text], return_tensors="pt")
+        encoded.to(zh_en_model.device)
+        sequences = zh_en_model.generate(**encoded)
+        return zh_en_tokenizer.batch_decode(sequences, skip_special_tokens=True)[0]
+
+# input = "青春不能回头，所以青春没有终点。 ——《火影忍者》"
+# print(input, translate(input))
+
+
+def text_generate(text_pipe,input,seed=None):
+    
+    if seed==None:
+        seed = random.randint(100, 1000000)
+    
+    set_seed(seed)
+
+    for count in range(6):    
+        sequences = text_pipe(input, max_length=random.randint(60, 90), num_return_sequences=8)
+        list = []
+        for sequence in sequences:
+            line = sequence['generated_text'].strip()
+            if line != input and len(line) > (len(input) + 4) and line.endswith((":", "-", "—")) is False:
+                list.append(line)
+
+        result = "\n".join(list)
+        result = re.sub('[^ ]+\.[^ ]+','', result)
+        result = result.replace("<", "").replace(">", "")
+        if result != "":
+            return result
+        if count == 5:
+            return result
+
+# input = "Youth can't turn back, so there's no end to youth."
+# print(input, text_generate(input))
+
+    
+import re
+
+def correct_prompt_syntax(prompt=""):
+
+    # print("input prompt",prompt)
+    corrected_elements = []
+    # 处理成统一的英文标点
+    prompt = prompt.replace('（', '(').replace('）', ')').replace('，', ',').replace(';', ',').replace('。', '.').replace('：',':')
+    # 删除多余的空格
+    prompt = re.sub(r'\s+', ' ', prompt).strip()
+    prompt = prompt.replace("< ","<").replace(" >",">").replace("( ","(").replace(" )",")").replace("[ ","[").replace(' ]',']')
+
+    # 分词
+    prompt_elements = prompt.split(',')
+
+    def balance_brackets(element, open_bracket, close_bracket):
+        open_brackets_count = element.count(open_bracket)
+        close_brackets_count = element.count(close_bracket)
+        return element + close_bracket * (open_brackets_count - close_brackets_count)
+
+    for element in prompt_elements:
+        element = element.strip()
+
+        # 处理空元素
+        if not element:
+            continue
+
+        # 检查并处理圆括号、方括号、尖括号
+        if element[0] in '([':
+            corrected_element = balance_brackets(element, '(', ')') if element[0] == '(' else balance_brackets(element, '[', ']')
+        elif element[0] == '<':
+            corrected_element = balance_brackets(element, '<', '>')
+        else:
+            # 删除开头的右括号或右方括号
+            corrected_element = element.lstrip(')]')
+
+        corrected_elements.append(corrected_element)
+
+    # 重组修正后的prompt
+    return  ','.join(corrected_elements)
+
+
+# # 示例使用
+# test_prompt = "((middle-century castles)), [forsaken: 0.8], (mystery dragons: 1.3, mist forests, sunsets, quiet; (((dummy)), [fisting city: 0.5] background, radiant, soft and flavoured,] promising mountains, ((starry: 1.6), [[crowds], [middle-century castle: urban landscapes of the future: 0.5], [yellow: bright sun: 0.7], overlooking"
+# corrected_prompt = correct_prompt_syntax(test_prompt)
+# print(corrected_prompt)
+
+def detect_language(input_str):
+    # 统计中文和英文字符的数量
+    count_cn = count_en = 0
+    for char in input_str:
+        if '\u4e00' <= char <= '\u9fff':
+            count_cn += 1
+        elif char.isalpha():
+            count_en += 1
+
+    # 根据统计的字符数量判断主要语言
+    if count_cn > count_en:
+        return "cn"
+    elif count_en > count_cn:
+        return "en"
+    else:
+        return "unknow"
+
+
+
+
+#定义Prompt文法
+grammar = """
+start: sentence
+sentence: phrase ("," phrase)*
+phrase: emphasis | weight | word | lora | embedding | schedule
+emphasis: "(" sentence ")" -> emphasis
+        | "[" sentence "]" -> weak_emphasis
+weight: "(" word ":" NUMBER ")"
+schedule: "[" word ":" word ":" NUMBER "]"
+lora: "<" WORD ":" WORD (":" NUMBER)? (":" NUMBER)? ">"
+embedding: "embedding" ":" WORD (":" NUMBER)? (":" NUMBER)?
+word: WORD
+
+NUMBER: /\s*-?\d+(\.\d+)?\s*/
+WORD: /[^,:\(\)\[\]<>]+/
+"""
+
+
+
+@v_args(inline=True)  # Decorator to flatten the tree directly into the function arguments
+class ChinesePromptTranslate(Transformer):
+ 
+    def sentence(self, *args):
+        return ", ".join(args)
+
+    def phrase(self, *args):
+        return "".join(args)
+
+    def emphasis(self, *args):
+        # Reconstruct the emphasis with translated content
+        return "(" + "".join(args) + ")"
+
+    def weak_emphasis(self, *args):
+        print('weak_emphasis:',args)
+        return "[" + "".join(args) + "]"
+    
+    def embedding(self,*args):
+        print('prompt embedding',args[0])
+        if len(args) == 1:
+            # print('prompt embedding',str(args[0]))
+            # 只传递了一个参数，意味着只有embedding名称没有数字
+            embedding_name = str(args[0])
+            return f"embedding:{embedding_name}"
+        elif len(args) > 1:
+            embedding_name,*numbers = args
+           
+            if len(numbers)==2:
+                return f"embedding:{embedding_name}:{numbers[0]}:{numbers[1]}"
+            elif len(numbers)==1:
+                return f"embedding:{embedding_name}:{numbers[0]}"
+            else:
+                return f"embedding:{embedding_name}"
+
+    def lora(self,*args):
+        print('lora prompt',*args)
+        if len(args) == 1:
+            return f"<lora:{loar_name}>"
+        elif len(args) > 1: 
+            # print('lora', args)
+            _,loar_name,*numbers = args
+            loar_name = str(loar_name).strip()
+            if len(numbers)==2:
+                return f"<lora:{loar_name}:{numbers[0]}:{numbers[1]}>"
+            elif len(numbers)==1:
+                return f"<lora:{loar_name}:{numbers[0]}>"
+            else:
+                return f"<lora:{loar_name}>"
+    
+    def weight(self, word,number):
+        translated_word = translate(str(word)).rstrip('.')
+        return f"({translated_word}:{str(number).strip()})"
+    
+    def schedule(self,*args):
+        print('prompt schedule',args)
+        data = [str(arg).strip() for arg in args] 
+        
+        return f"[{':'.join(data)}]"
+
+    def word(self, word):
+        # Translate each word using the dictionary
+        if detect_language(str(word)) == "cn":
+            return  translate(str(word)).rstrip('.')
+        else:
+            return str(word).rstrip('.')
+        
+class ChinesePrompt:
+
+    global _available
+    available=_available
+    
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+               "text": ("STRING",{"multiline": True,"default": "", "dynamicPrompts": False}),
+               "generation": (["on","off"],{"default": "off"}), 
+                             },
+               
+                "optional":{
+                    "seed":("INT", {"default": 100, "min": 100, "max":  0xffffffffffffffff}), 
+                    
+                },
+
+                }
+    
+    RETURN_TYPES = ("STRING",)
+    RETURN_NAMES = ("prompt",)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Prompt"
+    OUTPUT_NODE = True
+    INPUT_IS_LIST = True
+    OUTPUT_IS_LIST = (True,)
+
+    global text_pipe,zh_en_model,zh_en_tokenizer
+
+    text_pipe= None
+    zh_en_model=None
+    zh_en_tokenizer=None 
+
+    def run(self,text,seed,generation):
+        
+
+        seed=seed[0]
+        generation=generation[0]
+
+        # 进度条
+        pbar = comfy.utils.ProgressBar(len(text)+1)
+        texts = [correct_prompt_syntax(t) for t in text]
+
+        global text_pipe,zh_en_model,zh_en_tokenizer
+        if zh_en_model==None:
+            zh_en_model = AutoModelForSeq2SeqLM.from_pretrained(zh_en_model_path).eval()
+            zh_en_tokenizer = AutoTokenizer.from_pretrained(zh_en_model_path,padding=True, truncation=True)
+        
+        zh_en_model.to("cuda" if torch.cuda.is_available() else "cpu")
+            # zh_en_tokenizer.to("cuda" if torch.cuda.is_available() else "cpu")
+        
+        text_pipe=pipeline('text-generation', model=text_generator_model_path,device="cuda" if torch.cuda.is_available() else "cpu")
+        
+        # text_pipe.model.to("cuda" if torch.cuda.is_available() else "cpu")
+
+        prompt_result=[]
+
+        # print('zh_en_model device',zh_en_model.device,text_pipe.model.device,torch.cuda.current_device() )
+        en_texts=[]
+
+        for t in texts:
+            if t:
+                parser = Lark(grammar, start="start", parser="lalr", transformer=ChinesePromptTranslate())
+                try:
+                    result = parser.parse(t).children
+                    en_texts.append(result[0])
+                except:
+                    print(f"Error parsing '{t}'")
+                    t = translate(str(t))
+                    en_texts.append(t)
+                
+
+        zh_en_model.to('cpu')
+        print("test en_text",en_texts)
+        # en_text.to("cuda" if torch.cuda.is_available() else "cpu")
+
+        pbar.update(1)
+        for t in en_texts:
+            if generation=='on':
+                prompt =text_generate(text_pipe,t,seed)
+                # 多条，还是单条
+                lines = prompt.split("\n")
+                longest_line = max(lines, key=len)
+                # print(longest_line)
+                prompt_result.append(longest_line)
+            else:
+                prompt_result.append(t)
+            pbar.update(1)
+
+        text_pipe.model.to('cpu')
+
+        print('prompt_result',prompt_result,)
+        # prompt_result = [','.join(correct_prompt_syntax(p)) for p in prompt_result]
+        if len(prompt_result)==0:
+            prompt_result=[""]
+        return {
+            "ui":{
+                    "prompt": prompt_result
+                },
+            "result": (prompt_result,)}
+
+
+
+
+class PromptGenerate:
+
+    global _available
+    available=_available
+    
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "text": ("STRING",{"multiline": True,"default": "", "dynamicPrompts": False}),
+                             },
+
+                "optional":{
+                    "multiple": (["off","on"],), 
+                    "seed":("INT", {"default": 100, "min": 100, "max":  0xffffffffffffffff}), 
+                },
+
+                }
+    
+    RETURN_TYPES = ("STRING",)
+    RETURN_NAMES = ("prompt",)
+
+    FUNCTION = "run"
+
+    CATEGORY = "♾️Mixlab/Prompt"
+    OUTPUT_NODE = True
+    INPUT_IS_LIST = True
+    OUTPUT_IS_LIST = (True,)
+
+    global text_pipe
+
+    text_pipe= None
+    # 
+
+    def run(self,text,multiple,seed):
+        global text_pipe
+
+        seed=seed[0]
+
+        multiple=multiple[0]
+
+        # 进度条
+        pbar = comfy.utils.ProgressBar(len(text))
+        
+        text_pipe=pipeline('text-generation', model=text_generator_model_path,device="cuda" if torch.cuda.is_available() else "cpu")
+        
+        prompt_result=[]
+
+        for t in text:
+            prompt =text_generate(text_pipe,t,seed)
+            prompt = prompt.split("\n")
+            if multiple=='off':
+                prompt = [max(prompt, key=len)]
+            
+            for p in prompt:
+                prompt_result.append(p)
+            pbar.update(1)
+
+        text_pipe.model.to('cpu')
+
+        return {
+            "ui":{
+                    "prompt": prompt_result
+                },
+            "result": (prompt_result,)}
@@ -0,0 +1,180 @@
+import sys
+from os import path
+sys.path.insert(0, path.dirname(__file__))
+from PIL import Image
+import numpy as np
+import torch
+
+from folder_paths import get_folder_paths, get_full_path, get_save_image_path, get_output_directory,models_dir
+from comfy.model_management import get_torch_device
+from .tsr.system import TSR
+
+import comfy.utils
+
+
+def get_triposr_model_path():
+    try:
+        return path.join(get_folder_paths('triposr')[0],'model.ckpt')
+    except:
+        return path.join(path.join(models_dir, "triposr"),'model.ckpt')
+    
+triposr_model_path=get_triposr_model_path()
+
+
+# Tensor to PIL
+def tensor2pil(image):
+    return Image.fromarray(np.clip(255. * image.cpu().numpy().squeeze(), 0, 255).astype(np.uint8))
+
+# Convert PIL to Tensor
+def pil2tensor(image):
+    return torch.from_numpy(np.array(image).astype(np.float32) / 255.0).unsqueeze(0)
+
+
+def fill_background(image):
+    im = np.array(image).astype(np.float32) / 255.0
+    im = im[:, :, :3] * im[:, :, 3:4] + (1 - im[:, :, 3:4]) * 0.5
+    im = Image.fromarray((im * 255.0).astype(np.uint8))
+    return im
+
+
+class LoadTripoSRModel:
+    def __init__(self):
+        self.initialized_model = None
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                # "model": (get_filename_list("checkpoints"),),
+                "chunk_size": ("INT", {"default": 8192, "min": 0, "max": 10000})
+            }
+        }
+
+    RETURN_TYPES = ("TRIPOSR_MODEL",)
+    FUNCTION = "run"
+    CATEGORY = "♾️Mixlab/3D/TripoSR"
+
+    def run(self, chunk_size):
+        device = get_torch_device()
+
+        if not torch.cuda.is_available():
+            device = "cpu"
+
+        if not self.initialized_model:
+            # triposr_model_path
+            print("#Loading TripoSR model",triposr_model_path)
+            self.initialized_model = TSR.from_pretrained_custom(
+                weight_path=triposr_model_path,
+                config_path=path.join(path.dirname(__file__), "tsr/config.yaml")
+            )
+            self.initialized_model.renderer.set_chunk_size(chunk_size)
+            self.initialized_model.to(device)
+
+        return (self.initialized_model,)
+
+
+class TripoSRSampler:
+    def __init__(self):
+        self.initialized_model = None
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "model": ("TRIPOSR_MODEL",),
+                "image": ("IMAGE",),
+                "resolution": ("INT", {"default": 256, "min": 128, "max": 12288}),
+                "threshold": ("FLOAT", {"default": 25.0, "min": 0.0, "step": 0.01}),
+                "device":(["auto","cpu"],), 
+            },
+            "optional": {
+                "mask": ("MASK",)
+            }
+        }
+
+    RETURN_TYPES = ("MESH",)
+    FUNCTION = "run"
+    CATEGORY = "♾️Mixlab/3D/TripoSR"
+
+    def run(self, model, image, resolution, threshold,device='auto', mask=None):
+
+        reference_image=image
+        reference_mask=mask
+
+        device = get_torch_device()
+
+        if not torch.cuda.is_available():
+            device = "cpu"
+
+        if device=='cpu':
+            device = "cpu"
+
+        print('#TripoSRSampler device',device)
+
+        to_images=[]
+
+        for i in range(len(reference_image)):
+
+            image = reference_image[i]
+            
+            if reference_mask is not None:
+                mask = reference_mask[i].unsqueeze(2)
+                image = torch.cat((image, mask), dim=2).detach().cpu().numpy()
+                image = Image.fromarray(np.clip(255. * image, 0, 255).astype(np.uint8))
+                image = fill_background(image)
+            else:
+                image = tensor2pil(image)
+
+            image = image.convert('RGB')
+
+            to_images.append(image)
+
+        # 进度条
+        pbar = comfy.utils.ProgressBar(len(to_images))
+        def callback(c):
+            pbar.update(1)    
+
+        scene_codes = model(to_images, device)
+        meshes = model.extract_mesh(scene_codes, resolution=resolution, threshold=threshold,callback=callback)
+
+        del model
+        return (meshes,)
+
+
+class SaveTripoSRMesh:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "mesh": ("MESH",),
+                # "format":(["glb","obj"],), 
+                "filename_prefix":("STRING", {"multiline": False,"default": "TripoSR_"})
+            }
+        }
+
+    RETURN_TYPES = ()
+    OUTPUT_NODE = True
+    FUNCTION = "run"
+    CATEGORY = "♾️Mixlab/3D/TripoSR"
+
+    def run(self, mesh,filename_prefix):
+        format='glb'
+        saved = list()
+        full_output_folder, filename, counter, subfolder, filename_prefix = get_save_image_path(filename_prefix,
+                                                                                                get_output_directory())
+
+        for (index, single_mesh) in enumerate(mesh):
+            filename_with_batch_num = filename.replace("%batch_num%", str(index))
+            file = f"{filename_with_batch_num}_{counter:05}_.{format}"
+            single_mesh.apply_transform(np.array([[1, 0, 0, 0], [0, 0, 1, 0], [0, -1, 0, 0], [0, 0, 0, 1]]))
+            single_mesh.export(path.join(full_output_folder, file))
+            saved.append({
+                "filename": file,
+                "type": "output",
+                "subfolder": subfolder
+            })
+
+        return {"ui": {"mesh": saved}}
+
+
+
@@ -1,179 +0,0 @@
-# https://github.com/openai/consistencydecoder/blob/main/consistencydecoder/__init__.py
-
-import folder_paths
-from comfy import model_management
-
-import math
-import torch
-import numpy as np
-from PIL import Image 
-
-
-class ConsistencyDecoderWrapper:
-    def __init__(self, decoder):
-        self.decoder = decoder
-    def decode(self, x):
-        return self.decoder(x)
-
-def _extract_into_tensor(arr, timesteps, broadcast_shape):
-    
-    # from: https://github.com/openai/guided-diffusion/blob/22e0df8183507e13a7813f8d38d51b072ca1e67c/guided_diffusion/gaussian_diffusion.py#L895    """
-    res = arr[timesteps].float()
-    dims_to_append = len(broadcast_shape) - len(res.shape)
-    return res[(...,) + (None,) * dims_to_append]
-
-
-def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
-    # from: https://github.com/openai/guided-diffusion/blob/22e0df8183507e13a7813f8d38d51b072ca1e67c/guided_diffusion/gaussian_diffusion.py#L45
-    betas = []
-    for i in range(num_diffusion_timesteps):
-        t1 = i / num_diffusion_timesteps
-        t2 = (i + 1) / num_diffusion_timesteps
-        betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
-    return torch.tensor(betas)
-
-class ConsistencyDecoder:
-    def __init__(self, device="cuda:0", download_target=""):
-        self.n_distilled_steps = 64
-        # download_target = _download("https://openaipublic.azureedge.net/diff-vae/c9cebd3132dd9c42936d803e33424145a748843c8f716c0814838bdc8a2fe7cb/decoder.pt", download_root)
-        self.ckpt = torch.jit.load(download_target).to(device)
-        self.device = device 
-        sigma_data = 0.5
-        betas = betas_for_alpha_bar(
-            1024, lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
-        ).to(device)
-        alphas = 1.0 - betas
-        alphas_cumprod = torch.cumprod(alphas, dim=0)
-        self.sqrt_alphas_cumprod = torch.sqrt(alphas_cumprod)
-        self.sqrt_one_minus_alphas_cumprod = torch.sqrt(1.0 - alphas_cumprod)
-        sqrt_recip_alphas_cumprod = torch.sqrt(1.0 / alphas_cumprod)
-        sigmas = torch.sqrt(1.0 / alphas_cumprod - 1)
-        self.c_skip = (
-            sqrt_recip_alphas_cumprod
-            * sigma_data**2
-            / (sigmas**2 + sigma_data**2)
-        )
-        self.c_out = sigmas * sigma_data / (sigmas**2 + sigma_data**2) ** 0.5
-        self.c_in = sqrt_recip_alphas_cumprod / (sigmas**2 + sigma_data**2) ** 0.5
-
-    @staticmethod
-    def round_timesteps(
-        timesteps, total_timesteps, n_distilled_steps, truncate_start=True
-    ):
-        with torch.no_grad():
-            space = torch.div(total_timesteps, n_distilled_steps, rounding_mode="floor")
-            rounded_timesteps = (
-                torch.div(timesteps, space, rounding_mode="floor") + 1
-            ) * space
-            if truncate_start:
-                rounded_timesteps[rounded_timesteps == total_timesteps] -= space
-            else:
-                rounded_timesteps[rounded_timesteps == total_timesteps] -= space
-                rounded_timesteps[rounded_timesteps == 0] += space
-            return rounded_timesteps
-
-    @staticmethod
-    def ldm_transform_latent(z, extra_scale_factor=1):
-        channel_means = [0.38862467, 0.02253063, 0.07381133, -0.0171294]
-        channel_stds = [0.9654121, 1.0440036, 0.76147926, 0.77022034]
-
-        if len(z.shape) != 4:
-            raise ValueError()
-
-        z = z * 0.18215
-        channels = [z[:, i] for i in range(z.shape[1])]
-
-        channels = [
-            extra_scale_factor * (c - channel_means[i]) / channel_stds[i]
-            for i, c in enumerate(channels)
-        ]
-        return torch.stack(channels, dim=1)
-
-    @torch.no_grad()
-    def __call__(
-        self,
-        features: torch.Tensor,
-        schedule=[1.0, 0.5],
-    ):
-        features = self.ldm_transform_latent(features)
-  
-        ts = self.round_timesteps(
-            torch.arange(0, 1024),
-            1024,
-            self.n_distilled_steps,
-            truncate_start=False,
-        )
-        shape = (
-            features.size(0),
-            3,
-            8 * features.size(2),
-            8 * features.size(3),
-        )
-        
-        x_start = torch.zeros(shape, device=features.device, dtype=features.dtype)
-        schedule_timesteps = [int((1024 - 1) * s) for s in schedule]
-        for i in schedule_timesteps:
-            t = ts[i].item()
-            t_ = torch.tensor([t] * features.shape[0]).to(self.device)
-            noise = torch.randn_like(x_start)
-            
-            x_start = (
-                _extract_into_tensor(self.sqrt_alphas_cumprod, t_, x_start.shape)
-                * x_start
-                + _extract_into_tensor(
-                    self.sqrt_one_minus_alphas_cumprod, t_, x_start.shape
-                )
-                * noise
-            )
-            c_in = _extract_into_tensor(self.c_in, t_, x_start.shape)
-            model_output = self.ckpt(c_in * x_start, t_, features=features)
-            B, C = x_start.shape[:2]
-            model_output, _ = torch.split(model_output, C, dim=1)
-            pred_xstart = (
-                _extract_into_tensor(self.c_out, t_, x_start.shape) * model_output
-                + _extract_into_tensor(self.c_skip, t_, x_start.shape) * x_start
-            ).clamp(-1, 1)
-            x_start = pred_xstart
-        return x_start
-
-
-
-class VAELoader:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {"required": { "vae_name": (folder_paths.get_filename_list("vae"), )}}
-    RETURN_TYPES = ("VAE",)
-    FUNCTION = "load_vae"
-
-    CATEGORY = "♾️Mixlab/ConsistencyDecoder"
-
-    #TODO: scale factor?
-    def load_vae(self, vae_name):
-        vae_path = folder_paths.get_full_path("vae", vae_name)
-        device = 'cuda:0'
-        # print('device',device)
-        consistencyDecoder = ConsistencyDecoder(device=device,
-                                 download_target=vae_path) # Model size: 2.49 GB
-        vae = ConsistencyDecoderWrapper(consistencyDecoder)
-        return (vae,)
-    
-
-class VAEDecode:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {"required": { "samples": ("LATENT", ), "vae": ("VAE", )}}
-    RETURN_TYPES = ("IMAGE",)
-    FUNCTION = "decode"
-
-    CATEGORY = "♾️Mixlab/ConsistencyDecoder"
-
-    def decode(self, vae, samples):
-        image = vae.decode(samples["samples"].to("cuda:0"))
-        image = image[0].cpu().numpy()
-        image = (image + 1.0) * 127.5
-        image = image.clip(0, 255).astype(np.uint8)
-        image = Image.fromarray(image.transpose(1, 2, 0))
-        image = image.convert("RGB")
-        image = np.array(image).astype(np.float32) / 255.0
-        image = torch.from_numpy(image)[None,]
-        return (image, )
@@ -0,0 +1,45 @@
+from comfy.ldm.modules.attention import default, optimized_attention, optimized_attention_masked
+from .style_functions import adain, concat_first
+
+class VisualStyleProcessor(object):
+    def __init__(self, 
+        module_self, 
+        keys_scale: float = 1.0,
+        enabled: bool = True, 
+        adain_queries: bool = True,
+        adain_keys: bool = True,
+        adain_values: bool = False 
+    ):
+        self.module_self = module_self
+        self.keys_scale = keys_scale
+        self.enabled = enabled
+        self.adain_queries = adain_queries
+        self.adain_keys = adain_keys
+        self.adain_values = adain_values
+
+    def visual_style_forward(self, x, context, value, mask=None):
+        q = self.module_self.to_q(x)
+        context = default(context, x)
+        k = self.module_self.to_k(context)
+        if value is not None:
+            v = self.module_self.to_v(value)
+            del value
+        else:
+            v = self.module_self.to_v(context)
+
+        if self.enabled:
+            if self.adain_queries:
+                q = adain(q)
+            if self.adain_keys:
+                k = adain(k)
+            if self.adain_values:
+                v = adain(v)
+            
+            k = concat_first(k, -2, self.keys_scale)
+            v = concat_first(v, -2)
+
+        if mask is None:
+            out = optimized_attention(q, k, v, self.module_self.heads)
+        else:
+            out = optimized_attention_masked(q, k, v, self.module_self.heads, mask)
+        return self.module_self.to_out(out)
@@ -0,0 +1,60 @@
+import torch
+
+from einops import rearrange
+from dataclasses import dataclass
+
+T = torch.Tensor
+
+@dataclass(frozen=True)
+class StyleAlignedArgs:
+    share_group_norm: bool = True
+    share_layer_norm: bool = True,
+    share_attention: bool = True
+    adain_queries: bool = True
+    adain_keys: bool = True
+    adain_values: bool = False
+    full_attention_share: bool = False
+    keys_scale: float = 1.
+    only_self_level: float = 0.
+
+def expand_first(feat: T, scale=1., ) -> T:
+    b = feat.shape[0]
+    feat_style = torch.stack((feat[0], feat[b // 2])).unsqueeze(1)
+    if scale == 1:
+        feat_style = feat_style.expand(2, b // 2, *feat.shape[1:])
+    else:
+        feat_style = feat_style.repeat(1, b // 2, 1, 1, 1)
+        feat_style = torch.cat([feat_style[:, :1], scale * feat_style[:, 1:]], dim=1)
+    return feat_style.reshape(*feat.shape)
+
+
+def concat_first(feat: T, dim=2, scale=1.) -> T:
+    feat_style = expand_first(feat, scale=scale)
+    return torch.cat((feat, feat_style), dim=dim)
+
+
+def calc_mean_std(feat, eps: float = 1e-5) -> tuple[T, T]:
+    feat_std = (feat.var(dim=-2, keepdims=True) + eps).sqrt()
+    feat_mean = feat.mean(dim=-2, keepdims=True)
+    return feat_mean, feat_std
+
+
+def adain(feat: T) -> T:
+    feat_mean, feat_std = calc_mean_std(feat)
+    feat_style_mean = expand_first(feat_mean)
+    feat_style_std = expand_first(feat_std)
+    feat = (feat - feat_mean) / feat_std
+    feat = feat * feat_style_std + feat_style_mean
+    return feat
+
+def swapping_attention(key, value, chunk_size=2):
+    chunk_length = key.size()[0] // chunk_size  # [text-condition, null-condition]
+    reference_image_index = [0] * chunk_length  # [0 0 0 0 0]
+    key = rearrange(key, "(b f) d c -> b f d c", f=chunk_length)
+    key = key[:, reference_image_index]  # ref to all
+    key = rearrange(key, "b f d c -> (b f) d c")
+    value = rearrange(value, "(b f) d c -> b f d c", f=chunk_length)
+    value = value[:, reference_image_index]  # ref to all
+    value = rearrange(value, "b f d c -> (b f) d c")
+
+    return key, value
@@ -0,0 +1,172 @@
+import torch
+from PIL import Image, ImageOps, ImageSequence, ImageFile
+from PIL.PngImagePlugin import PngInfo
+
+import numpy as np
+import os
+import folder_paths 
+import node_helpers
+import hashlib
+
+# Tensor to PIL
+def tensor2pil(image):
+    return Image.fromarray(np.clip(255. * image.cpu().numpy().squeeze(), 0, 255).astype(np.uint8))
+
+# tensor 取hash值
+def tensor_to_hash(tensor):
+    # 将 Tensor 转换为 NumPy 数组
+    np_array = tensor.cpu().numpy()
+    
+    # 将 NumPy 数组转换为字节数据
+    byte_data = np_array.tobytes()
+    
+    # 计算哈希值
+    hash_value = hashlib.md5(byte_data).hexdigest()
+    
+    return hash_value
+
+
+def create_temp_file(image):
+    output_dir = folder_paths.get_temp_directory()
+
+    (
+            full_output_folder,
+            filename,
+            counter,
+            subfolder,
+            _,
+        ) = folder_paths.get_save_image_path('material', output_dir)
+
+    
+    image=tensor2pil(image)
+ 
+    image_file = f"{filename}_{counter:05}.png"
+     
+    image_path=os.path.join(full_output_folder, image_file)
+
+    image.save(image_path,compress_level=4)
+
+    return (image_path,[{
+    "filename": image_file,
+    "subfolder": subfolder,
+    "type": "temp"
+    }])
+
+
+# image - tensor - 文件路径
+# loadImage的方法（ 文件路径 - image-mask ）
+class EditMask:
+
+    def __init__(self):
+        self.image_id = None
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required":
+                    {"image": ("IMAGE",), # 表示一个张量
+                     
+                     },
+
+                    "optional":{
+                            "image_update": ("IMAGE_FILE",)
+                        },
+                   
+                }
+
+    CATEGORY = "♾️Mixlab/Mask"
+
+    RETURN_TYPES = ("IMAGE", "MASK")
+    RETURN_NAMES = ("image", "mask")
+
+    FUNCTION = "edit"
+
+    OUTPUT_NODE = True
+
+    def edit(self, image,image_update=None):
+
+        # 根据image输入来判断是否是新的图片
+        if self.image_id==None:
+            self.image_id=tensor_to_hash(image)
+            image_update=None
+        else:
+            image_id=tensor_to_hash(image)
+            if image_id!=self.image_id:
+                image_update=None
+                self.image_id=image_id
+
+
+        image_path=None
+        # print('#image_update',self.image_id,image_update)
+        if image_update==None:
+            print('--')
+        else:
+            if 'images' in image_update:
+                images=image_update['images']
+                filename=images[0]['filename']
+                subfolder=images[0]['subfolder']
+                type=images[0]['type']
+                name, base_dir=folder_paths.annotated_filepath(filename)
+                if type.endswith("output"):
+                    base_dir = folder_paths.get_output_directory() 
+                elif type.endswith("input"):
+                    base_dir = folder_paths.get_input_directory() 
+                elif type.endswith("temp"):
+                    base_dir = folder_paths.get_temp_directory() 
+                #base_dir = folder_paths.get_input_directory()  
+                # print(base_dir,subfolder, name)
+                image_path = os.path.join(base_dir,subfolder, name)
+        
+        if image_path==None:
+            image_path,images=create_temp_file(image)
+
+        print('#image_path',os.path.exists(image_path),image_path)
+        # image_path = folder_paths.get_annotated_filepath(image) #文件名
+        
+        if not os.path.exists(image_path):
+            image_path,images=create_temp_file(image)
+
+
+        img = node_helpers.pillow(Image.open, image_path)
+        
+        output_images = []
+        output_masks = []
+        w, h = None, None
+
+        excluded_formats = ['MPO']
+        
+        for i in ImageSequence.Iterator(img):
+            i = node_helpers.pillow(ImageOps.exif_transpose, i)
+
+            if i.mode == 'I':
+                i = i.point(lambda i: i * (1 / 255))
+            image = i.convert("RGB")
+
+            if len(output_images) == 0:
+                w = image.size[0]
+                h = image.size[1]
+            
+            if image.size[0] != w or image.size[1] != h:
+                continue
+            
+            image = np.array(image).astype(np.float32) / 255.0
+            image = torch.from_numpy(image)[None,]
+            if 'A' in i.getbands():
+                mask = np.array(i.getchannel('A')).astype(np.float32) / 255.0
+                mask = 1. - torch.from_numpy(mask)
+            else:
+                # 尺寸不对，需要按照image来
+                mask = torch.zeros((h, w), dtype=torch.float32, device="cpu")
+
+            output_images.append(image)
+            output_masks.append(mask.unsqueeze(0))
+
+        if len(output_images) > 1 and img.format not in excluded_formats:
+            output_image = torch.cat(output_images, dim=0)
+            output_mask = torch.cat(output_masks, dim=0)
+        else:
+            output_image = output_images[0]
+            output_mask = output_masks[0]
+
+        return {"ui":{"images": images},"result": (output_image, output_mask)}
+
+        # return (output_image, output_mask)
@@ -0,0 +1,8 @@
+import folder_paths
+
+# 外挂一个文件，用来编写新的节点
+def run(v):
+
+    output_dir = folder_paths.get_temp_directory()
+
+    print('1323',v,output_dir)
@@ -0,0 +1,38 @@
+cond_image_size: 512
+
+image_tokenizer_cls: tsr.models.tokenizers.image.DINOSingleImageTokenizer
+image_tokenizer:
+  pretrained_model_name_or_path: "facebook/dino-vitb16"
+
+tokenizer_cls: tsr.models.tokenizers.triplane.Triplane1DTokenizer
+tokenizer:
+  plane_size: 32
+  num_channels: 1024
+
+backbone_cls: tsr.models.transformer.transformer_1d.Transformer1D
+backbone:
+  in_channels: ${tokenizer.num_channels}
+  num_attention_heads: 16
+  attention_head_dim: 64
+  num_layers: 16
+  cross_attention_dim: 768
+
+post_processor_cls: tsr.models.network_utils.TriplaneUpsampleNetwork
+post_processor:
+  in_channels: 1024
+  out_channels: 40
+
+decoder_cls: tsr.models.network_utils.NeRFMLP
+decoder:
+  in_channels: 120 # 3 * 40
+  n_neurons: 64
+  n_hidden_layers: 9
+  activation: silu
+
+renderer_cls: tsr.models.nerf_renderer.TriplaneNeRFRenderer
+renderer:
+  radius: 0.87 # slightly larger than 0.5 * sqrt(3)
+  feature_reduction: concat
+  density_activation: exp
+  density_bias: -1.0
+  num_samples_per_ray: 128
@@ -0,0 +1,51 @@
+from typing import Callable, Optional, Tuple
+
+import numpy as np
+import torch
+import torch.nn as nn
+from skimage import measure
+
+
+class IsosurfaceHelper(nn.Module):
+    points_range: Tuple[float, float] = (0, 1)
+
+    @property
+    def grid_vertices(self) -> torch.FloatTensor:
+        raise NotImplementedError
+
+
+class MarchingCubeHelper(IsosurfaceHelper):
+    def __init__(self, resolution: int) -> None:
+        super().__init__()
+        self.resolution = resolution
+        #self.mc_func: Callable = marching_cubes
+        self._grid_vertices: Optional[torch.FloatTensor] = None
+
+    @property
+    def grid_vertices(self) -> torch.FloatTensor:
+        if self._grid_vertices is None:
+            # keep the vertices on CPU so that we can support very large resolution
+            x, y, z = (
+                torch.linspace(*self.points_range, self.resolution),
+                torch.linspace(*self.points_range, self.resolution),
+                torch.linspace(*self.points_range, self.resolution),
+            )
+            x, y, z = torch.meshgrid(x, y, z, indexing="ij")
+            verts = torch.cat(
+                [x.reshape(-1, 1), y.reshape(-1, 1), z.reshape(-1, 1)], dim=-1
+            ).reshape(-1, 3)
+            self._grid_vertices = verts
+        return self._grid_vertices
+
+    def forward(
+        self,
+        level: torch.FloatTensor,
+    ) -> Tuple[torch.FloatTensor, torch.LongTensor]:
+        level = -level.view(self.resolution, self.resolution, self.resolution)
+        v_pos, t_pos_idx, _, __ = measure.marching_cubes((level.detach().cpu() if level.is_cuda else level.detach()).numpy(), 0.0) #self.mc_func(level.detach(), 0.0)
+        v_pos = torch.from_numpy(v_pos.copy()).type(torch.FloatTensor).to(level.device)
+        t_pos_idx = torch.from_numpy(t_pos_idx.copy()).type(torch.LongTensor).to(level.device)
+        v_pos = v_pos[..., [0, 1, 2]]
+        t_pos_idx = t_pos_idx[..., [1, 0, 2]]
+        v_pos = v_pos / (self.resolution - 1.0)
+        return v_pos, t_pos_idx
@@ -0,0 +1,180 @@
+from dataclasses import dataclass
+from typing import Dict
+
+import torch
+import torch.nn.functional as F
+from einops import rearrange, reduce
+
+from ..utils import (
+    BaseModule,
+    chunk_batch,
+    get_activation,
+    rays_intersect_bbox,
+    scale_tensor,
+)
+
+
+class TriplaneNeRFRenderer(BaseModule):
+    @dataclass
+    class Config(BaseModule.Config):
+        radius: float
+
+        feature_reduction: str = "concat"
+        density_activation: str = "trunc_exp"
+        density_bias: float = -1.0
+        color_activation: str = "sigmoid"
+        num_samples_per_ray: int = 128
+        randomized: bool = False
+
+    cfg: Config
+
+    def configure(self) -> None:
+        assert self.cfg.feature_reduction in ["concat", "mean"]
+        self.chunk_size = 0
+
+    def set_chunk_size(self, chunk_size: int):
+        assert (
+            chunk_size >= 0
+        ), "chunk_size must be a non-negative integer (0 for no chunking)."
+        self.chunk_size = chunk_size
+
+    def query_triplane(
+        self,
+        decoder: torch.nn.Module,
+        positions: torch.Tensor,
+        triplane: torch.Tensor,
+    ) -> Dict[str, torch.Tensor]:
+        input_shape = positions.shape[:-1]
+        positions = positions.view(-1, 3)
+
+        # positions in (-radius, radius)
+        # normalized to (-1, 1) for grid sample
+        positions = scale_tensor(
+            positions, (-self.cfg.radius, self.cfg.radius), (-1, 1)
+        )
+
+        def _query_chunk(x):
+            indices2D: torch.Tensor = torch.stack(
+                (x[..., [0, 1]], x[..., [0, 2]], x[..., [1, 2]]),
+                dim=-3,
+            )
+            out: torch.Tensor = F.grid_sample(
+                rearrange(triplane, "Np Cp Hp Wp -> Np Cp Hp Wp", Np=3),
+                rearrange(indices2D, "Np N Nd -> Np () N Nd", Np=3),
+                align_corners=False,
+                mode="bilinear",
+            )
+            if self.cfg.feature_reduction == "concat":
+                out = rearrange(out, "Np Cp () N -> N (Np Cp)", Np=3)
+            elif self.cfg.feature_reduction == "mean":
+                out = reduce(out, "Np Cp () N -> N Cp", Np=3, reduction="mean")
+            else:
+                raise NotImplementedError
+
+            net_out: Dict[str, torch.Tensor] = decoder(out)
+            return net_out
+
+        if self.chunk_size > 0:
+            net_out = chunk_batch(_query_chunk, self.chunk_size, positions)
+        else:
+            net_out = _query_chunk(positions)
+
+        net_out["density_act"] = get_activation(self.cfg.density_activation)(
+            net_out["density"] + self.cfg.density_bias
+        )
+        net_out["color"] = get_activation(self.cfg.color_activation)(
+            net_out["features"]
+        )
+
+        net_out = {k: v.view(*input_shape, -1) for k, v in net_out.items()}
+
+        return net_out
+
+    def _forward(
+        self,
+        decoder: torch.nn.Module,
+        triplane: torch.Tensor,
+        rays_o: torch.Tensor,
+        rays_d: torch.Tensor,
+        **kwargs,
+    ):
+        rays_shape = rays_o.shape[:-1]
+        rays_o = rays_o.view(-1, 3)
+        rays_d = rays_d.view(-1, 3)
+        n_rays = rays_o.shape[0]
+
+        t_near, t_far, rays_valid = rays_intersect_bbox(rays_o, rays_d, self.cfg.radius)
+        t_near, t_far = t_near[rays_valid], t_far[rays_valid]
+
+        t_vals = torch.linspace(
+            0, 1, self.cfg.num_samples_per_ray + 1, device=triplane.device
+        )
+        t_mid = (t_vals[:-1] + t_vals[1:]) / 2.0
+        z_vals = t_near * (1 - t_mid[None]) + t_far * t_mid[None]  # (N_rays, N_samples)
+
+        xyz = (
+            rays_o[:, None, :] + z_vals[..., None] * rays_d[..., None, :]
+        )  # (N_rays, N_sample, 3)
+
+        mlp_out = self.query_triplane(
+            decoder=decoder,
+            positions=xyz,
+            triplane=triplane,
+        )
+
+        eps = 1e-10
+        # deltas = z_vals[:, 1:] - z_vals[:, :-1] # (N_rays, N_samples)
+        deltas = t_vals[1:] - t_vals[:-1]  # (N_rays, N_samples)
+        alpha = 1 - torch.exp(
+            -deltas * mlp_out["density_act"][..., 0]
+        )  # (N_rays, N_samples)
+        accum_prod = torch.cat(
+            [
+                torch.ones_like(alpha[:, :1]),
+                torch.cumprod(1 - alpha[:, :-1] + eps, dim=-1),
+            ],
+            dim=-1,
+        )
+        weights = alpha * accum_prod  # (N_rays, N_samples)
+        comp_rgb_ = (weights[..., None] * mlp_out["color"]).sum(dim=-2)  # (N_rays, 3)
+        opacity_ = weights.sum(dim=-1)  # (N_rays)
+
+        comp_rgb = torch.zeros(
+            n_rays, 3, dtype=comp_rgb_.dtype, device=comp_rgb_.device
+        )
+        opacity = torch.zeros(n_rays, dtype=opacity_.dtype, device=opacity_.device)
+        comp_rgb[rays_valid] = comp_rgb_
+        opacity[rays_valid] = opacity_
+
+        comp_rgb += 1 - opacity[..., None]
+        comp_rgb = comp_rgb.view(*rays_shape, 3)
+
+        return comp_rgb
+
+    def forward(
+        self,
+        decoder: torch.nn.Module,
+        triplane: torch.Tensor,
+        rays_o: torch.Tensor,
+        rays_d: torch.Tensor,
+    ) -> Dict[str, torch.Tensor]:
+        if triplane.ndim == 4:
+            comp_rgb = self._forward(decoder, triplane, rays_o, rays_d)
+        else:
+            comp_rgb = torch.stack(
+                [
+                    self._forward(decoder, triplane[i], rays_o[i], rays_d[i])
+                    for i in range(triplane.shape[0])
+                ],
+                dim=0,
+            )
+
+        return comp_rgb
+
+    def train(self, mode=True):
+        self.randomized = mode and self.cfg.randomized
+        return super().train(mode=mode)
+
+    def eval(self):
+        self.randomized = False
+        return super().eval()
@@ -0,0 +1,124 @@
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+
+from ..utils import BaseModule
+
+
+class TriplaneUpsampleNetwork(BaseModule):
+    @dataclass
+    class Config(BaseModule.Config):
+        in_channels: int
+        out_channels: int
+
+    cfg: Config
+
+    def configure(self) -> None:
+        self.upsample = nn.ConvTranspose2d(
+            self.cfg.in_channels, self.cfg.out_channels, kernel_size=2, stride=2
+        )
+
+    def forward(self, triplanes: torch.Tensor) -> torch.Tensor:
+        triplanes_up = rearrange(
+            self.upsample(
+                rearrange(triplanes, "B Np Ci Hp Wp -> (B Np) Ci Hp Wp", Np=3)
+            ),
+            "(B Np) Co Hp Wp -> B Np Co Hp Wp",
+            Np=3,
+        )
+        return triplanes_up
+
+
+class NeRFMLP(BaseModule):
+    @dataclass
+    class Config(BaseModule.Config):
+        in_channels: int
+        n_neurons: int
+        n_hidden_layers: int
+        activation: str = "relu"
+        bias: bool = True
+        weight_init: Optional[str] = "kaiming_uniform"
+        bias_init: Optional[str] = None
+
+    cfg: Config
+
+    def configure(self) -> None:
+        layers = [
+            self.make_linear(
+                self.cfg.in_channels,
+                self.cfg.n_neurons,
+                bias=self.cfg.bias,
+                weight_init=self.cfg.weight_init,
+                bias_init=self.cfg.bias_init,
+            ),
+            self.make_activation(self.cfg.activation),
+        ]
+        for i in range(self.cfg.n_hidden_layers - 1):
+            layers += [
+                self.make_linear(
+                    self.cfg.n_neurons,
+                    self.cfg.n_neurons,
+                    bias=self.cfg.bias,
+                    weight_init=self.cfg.weight_init,
+                    bias_init=self.cfg.bias_init,
+                ),
+                self.make_activation(self.cfg.activation),
+            ]
+        layers += [
+            self.make_linear(
+                self.cfg.n_neurons,
+                4,  # density 1 + features 3
+                bias=self.cfg.bias,
+                weight_init=self.cfg.weight_init,
+                bias_init=self.cfg.bias_init,
+            )
+        ]
+        self.layers = nn.Sequential(*layers)
+
+    def make_linear(
+        self,
+        dim_in,
+        dim_out,
+        bias=True,
+        weight_init=None,
+        bias_init=None,
+    ):
+        layer = nn.Linear(dim_in, dim_out, bias=bias)
+
+        if weight_init is None:
+            pass
+        elif weight_init == "kaiming_uniform":
+            torch.nn.init.kaiming_uniform_(layer.weight, nonlinearity="relu")
+        else:
+            raise NotImplementedError
+
+        if bias:
+            if bias_init is None:
+                pass
+            elif bias_init == "zero":
+                torch.nn.init.zeros_(layer.bias)
+            else:
+                raise NotImplementedError
+
+        return layer
+
+    def make_activation(self, activation):
+        if activation == "relu":
+            return nn.ReLU(inplace=True)
+        elif activation == "silu":
+            return nn.SiLU(inplace=True)
+        else:
+            raise NotImplementedError
+
+    def forward(self, x):
+        inp_shape = x.shape[:-1]
+        x = x.reshape(-1, x.shape[-1])
+
+        features = self.layers(x)
+        features = features.reshape(*inp_shape, -1)
+        out = {"density": features[..., 0:1], "features": features[..., 1:4]}
+
+        return out
@@ -0,0 +1,72 @@
+from dataclasses import dataclass
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from huggingface_hub import hf_hub_download
+from transformers.models.vit.modeling_vit import ViTModel
+
+from ...utils import BaseModule
+import os
+import folder_paths
+model_path=os.path.join(folder_paths.models_dir,'triposr')
+
+
+class DINOSingleImageTokenizer(BaseModule):
+    @dataclass
+    class Config(BaseModule.Config):
+        pretrained_model_name_or_path: str = "facebook/dino-vitb16"
+        enable_gradient_checkpointing: bool = False
+
+    cfg: Config
+
+    def configure(self) -> None:
+        print('#Loading ViTModel:',os.path.join(model_path,self.cfg.pretrained_model_name_or_path))
+        self.model: ViTModel = ViTModel(
+            ViTModel.config_class.from_pretrained(
+                hf_hub_download(
+                    repo_id=self.cfg.pretrained_model_name_or_path,
+                    filename="config.json",
+                    local_dir=model_path,
+                    endpoint='https://hf-mirror.com'
+                )
+            )
+        )
+
+        if self.cfg.enable_gradient_checkpointing:
+            self.model.encoder.gradient_checkpointing = True
+
+        self.register_buffer(
+            "image_mean",
+            torch.as_tensor([0.485, 0.456, 0.406]).reshape(1, 1, 3, 1, 1),
+            persistent=False,
+        )
+        self.register_buffer(
+            "image_std",
+            torch.as_tensor([0.229, 0.224, 0.225]).reshape(1, 1, 3, 1, 1),
+            persistent=False,
+        )
+
+    def forward(self, images: torch.FloatTensor, **kwargs) -> torch.FloatTensor:
+        packed = False
+        if images.ndim == 4:
+            packed = True
+            images = images.unsqueeze(1)
+
+        batch_size, n_input_views = images.shape[:2]
+        images = (images - self.image_mean) / self.image_std
+        out = self.model(
+            rearrange(images, "B N C H W -> (B N) C H W"), interpolate_pos_encoding=True
+        )
+        local_features, global_features = out.last_hidden_state, out.pooler_output
+        local_features = local_features.permute(0, 2, 1)
+        local_features = rearrange(
+            local_features, "(B N) Ct Nt -> B N Ct Nt", B=batch_size
+        )
+        if packed:
+            local_features = local_features.squeeze(1)
+
+        return local_features
+
+    def detokenize(self, *args, **kwargs):
+        raise NotImplementedError
@@ -0,0 +1,45 @@
+import math
+from dataclasses import dataclass
+
+import torch
+import torch.nn as nn
+from einops import rearrange, repeat
+
+from ...utils import BaseModule
+
+
+class Triplane1DTokenizer(BaseModule):
+    @dataclass
+    class Config(BaseModule.Config):
+        plane_size: int
+        num_channels: int
+
+    cfg: Config
+
+    def configure(self) -> None:
+        self.embeddings = nn.Parameter(
+            torch.randn(
+                (3, self.cfg.num_channels, self.cfg.plane_size, self.cfg.plane_size),
+                dtype=torch.float32,
+            )
+            * 1
+            / math.sqrt(self.cfg.num_channels)
+        )
+
+    def forward(self, batch_size: int) -> torch.Tensor:
+        return rearrange(
+            repeat(self.embeddings, "Np Ct Hp Wp -> B Np Ct Hp Wp", B=batch_size),
+            "B Np Ct Hp Wp -> B Ct (Np Hp Wp)",
+        )
+
+    def detokenize(self, tokens: torch.Tensor) -> torch.Tensor:
+        batch_size, Ct, Nt = tokens.shape
+        assert Nt == self.cfg.plane_size**2 * 3
+        assert Ct == self.cfg.num_channels
+        return rearrange(
+            tokens,
+            "B Ct (Np Hp Wp) -> B Np Ct Hp Wp",
+            Np=3,
+            Hp=self.cfg.plane_size,
+            Wp=self.cfg.plane_size,
+        )
@@ -0,0 +1,653 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# --------
+#
+# Modified 2024 by the Tripo AI and Stability AI Team.
+#
+# Copyright (c) 2024 Tripo AI & Stability AI
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+class Attention(nn.Module):
+    r"""
+    A cross attention layer.
+
+    Parameters:
+        query_dim (`int`):
+            The number of channels in the query.
+        cross_attention_dim (`int`, *optional*):
+            The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`.
+        heads (`int`,  *optional*, defaults to 8):
+            The number of heads to use for multi-head attention.
+        dim_head (`int`,  *optional*, defaults to 64):
+            The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The dropout probability to use.
+        bias (`bool`, *optional*, defaults to False):
+            Set to `True` for the query, key, and value linear layers to contain a bias parameter.
+        upcast_attention (`bool`, *optional*, defaults to False):
+            Set to `True` to upcast the attention computation to `float32`.
+        upcast_softmax (`bool`, *optional*, defaults to False):
+            Set to `True` to upcast the softmax computation to `float32`.
+        cross_attention_norm (`str`, *optional*, defaults to `None`):
+            The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`.
+        cross_attention_norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups to use for the group norm in the cross attention.
+        added_kv_proj_dim (`int`, *optional*, defaults to `None`):
+            The number of channels to use for the added key and value projections. If `None`, no projection is used.
+        norm_num_groups (`int`, *optional*, defaults to `None`):
+            The number of groups to use for the group norm in the attention.
+        spatial_norm_dim (`int`, *optional*, defaults to `None`):
+            The number of channels to use for the spatial normalization.
+        out_bias (`bool`, *optional*, defaults to `True`):
+            Set to `True` to use a bias in the output linear layer.
+        scale_qk (`bool`, *optional*, defaults to `True`):
+            Set to `True` to scale the query and key by `1 / sqrt(dim_head)`.
+        only_cross_attention (`bool`, *optional*, defaults to `False`):
+            Set to `True` to only use cross attention and not added_kv_proj_dim. Can only be set to `True` if
+            `added_kv_proj_dim` is not `None`.
+        eps (`float`, *optional*, defaults to 1e-5):
+            An additional value added to the denominator in group normalization that is used for numerical stability.
+        rescale_output_factor (`float`, *optional*, defaults to 1.0):
+            A factor to rescale the output by dividing it with this value.
+        residual_connection (`bool`, *optional*, defaults to `False`):
+            Set to `True` to add the residual connection to the output.
+        _from_deprecated_attn_block (`bool`, *optional*, defaults to `False`):
+            Set to `True` if the attention block is loaded from a deprecated state dict.
+        processor (`AttnProcessor`, *optional*, defaults to `None`):
+            The attention processor to use. If `None`, defaults to `AttnProcessor2_0` if `torch 2.x` is used and
+            `AttnProcessor` otherwise.
+    """
+
+    def __init__(
+        self,
+        query_dim: int,
+        cross_attention_dim: Optional[int] = None,
+        heads: int = 8,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        bias: bool = False,
+        upcast_attention: bool = False,
+        upcast_softmax: bool = False,
+        cross_attention_norm: Optional[str] = None,
+        cross_attention_norm_num_groups: int = 32,
+        added_kv_proj_dim: Optional[int] = None,
+        norm_num_groups: Optional[int] = None,
+        out_bias: bool = True,
+        scale_qk: bool = True,
+        only_cross_attention: bool = False,
+        eps: float = 1e-5,
+        rescale_output_factor: float = 1.0,
+        residual_connection: bool = False,
+        _from_deprecated_attn_block: bool = False,
+        processor: Optional["AttnProcessor"] = None,
+        out_dim: int = None,
+    ):
+        super().__init__()
+        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.query_dim = query_dim
+        self.cross_attention_dim = (
+            cross_attention_dim if cross_attention_dim is not None else query_dim
+        )
+        self.upcast_attention = upcast_attention
+        self.upcast_softmax = upcast_softmax
+        self.rescale_output_factor = rescale_output_factor
+        self.residual_connection = residual_connection
+        self.dropout = dropout
+        self.fused_projections = False
+        self.out_dim = out_dim if out_dim is not None else query_dim
+
+        # we make use of this private variable to know whether this class is loaded
+        # with an deprecated state dict so that we can convert it on the fly
+        self._from_deprecated_attn_block = _from_deprecated_attn_block
+
+        self.scale_qk = scale_qk
+        self.scale = dim_head**-0.5 if self.scale_qk else 1.0
+
+        self.heads = out_dim // dim_head if out_dim is not None else heads
+        # for slice_size > 0 the attention score computation
+        # is split across the batch axis to save memory
+        # You can set slice_size with `set_attention_slice`
+        self.sliceable_head_dim = heads
+
+        self.added_kv_proj_dim = added_kv_proj_dim
+        self.only_cross_attention = only_cross_attention
+
+        if self.added_kv_proj_dim is None and self.only_cross_attention:
+            raise ValueError(
+                "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`."
+            )
+
+        if norm_num_groups is not None:
+            self.group_norm = nn.GroupNorm(
+                num_channels=query_dim, num_groups=norm_num_groups, eps=eps, affine=True
+            )
+        else:
+            self.group_norm = None
+
+        self.spatial_norm = None
+
+        if cross_attention_norm is None:
+            self.norm_cross = None
+        elif cross_attention_norm == "layer_norm":
+            self.norm_cross = nn.LayerNorm(self.cross_attention_dim)
+        elif cross_attention_norm == "group_norm":
+            if self.added_kv_proj_dim is not None:
+                # The given `encoder_hidden_states` are initially of shape
+                # (batch_size, seq_len, added_kv_proj_dim) before being projected
+                # to (batch_size, seq_len, cross_attention_dim). The norm is applied
+                # before the projection, so we need to use `added_kv_proj_dim` as
+                # the number of channels for the group norm.
+                norm_cross_num_channels = added_kv_proj_dim
+            else:
+                norm_cross_num_channels = self.cross_attention_dim
+
+            self.norm_cross = nn.GroupNorm(
+                num_channels=norm_cross_num_channels,
+                num_groups=cross_attention_norm_num_groups,
+                eps=1e-5,
+                affine=True,
+            )
+        else:
+            raise ValueError(
+                f"unknown cross_attention_norm: {cross_attention_norm}. Should be None, 'layer_norm' or 'group_norm'"
+            )
+
+        linear_cls = nn.Linear
+
+        self.linear_cls = linear_cls
+        self.to_q = linear_cls(query_dim, self.inner_dim, bias=bias)
+
+        if not self.only_cross_attention:
+            # only relevant for the `AddedKVProcessor` classes
+            self.to_k = linear_cls(self.cross_attention_dim, self.inner_dim, bias=bias)
+            self.to_v = linear_cls(self.cross_attention_dim, self.inner_dim, bias=bias)
+        else:
+            self.to_k = None
+            self.to_v = None
+
+        if self.added_kv_proj_dim is not None:
+            self.add_k_proj = linear_cls(added_kv_proj_dim, self.inner_dim)
+            self.add_v_proj = linear_cls(added_kv_proj_dim, self.inner_dim)
+
+        self.to_out = nn.ModuleList([])
+        self.to_out.append(linear_cls(self.inner_dim, self.out_dim, bias=out_bias))
+        self.to_out.append(nn.Dropout(dropout))
+
+        # set attention processor
+        # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
+        # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
+        # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
+        if processor is None:
+            processor = (
+                AttnProcessor2_0()
+                if hasattr(F, "scaled_dot_product_attention") and self.scale_qk
+                else AttnProcessor()
+            )
+        self.set_processor(processor)
+
+    def set_processor(self, processor: "AttnProcessor") -> None:
+        self.processor = processor
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        **cross_attention_kwargs,
+    ) -> torch.Tensor:
+        r"""
+        The forward method of the `Attention` class.
+
+        Args:
+            hidden_states (`torch.Tensor`):
+                The hidden states of the query.
+            encoder_hidden_states (`torch.Tensor`, *optional*):
+                The hidden states of the encoder.
+            attention_mask (`torch.Tensor`, *optional*):
+                The attention mask to use. If `None`, no mask is applied.
+            **cross_attention_kwargs:
+                Additional keyword arguments to pass along to the cross attention.
+
+        Returns:
+            `torch.Tensor`: The output of the attention layer.
+        """
+        # The `Attention` class can call different attention processors / attention functions
+        # here we simply pass along all tensors to the selected processor class
+        # For standard processors that are defined here, `**cross_attention_kwargs` is empty
+        return self.processor(
+            self,
+            hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+
+    def batch_to_head_dim(self, tensor: torch.Tensor) -> torch.Tensor:
+        r"""
+        Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size // heads, seq_len, dim * heads]`. `heads`
+        is the number of heads initialized while constructing the `Attention` class.
+
+        Args:
+            tensor (`torch.Tensor`): The tensor to reshape.
+
+        Returns:
+            `torch.Tensor`: The reshaped tensor.
+        """
+        head_size = self.heads
+        batch_size, seq_len, dim = tensor.shape
+        tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
+        tensor = tensor.permute(0, 2, 1, 3).reshape(
+            batch_size // head_size, seq_len, dim * head_size
+        )
+        return tensor
+
+    def head_to_batch_dim(self, tensor: torch.Tensor, out_dim: int = 3) -> torch.Tensor:
+        r"""
+        Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size, seq_len, heads, dim // heads]` `heads` is
+        the number of heads initialized while constructing the `Attention` class.
+
+        Args:
+            tensor (`torch.Tensor`): The tensor to reshape.
+            out_dim (`int`, *optional*, defaults to `3`): The output dimension of the tensor. If `3`, the tensor is
+                reshaped to `[batch_size * heads, seq_len, dim // heads]`.
+
+        Returns:
+            `torch.Tensor`: The reshaped tensor.
+        """
+        head_size = self.heads
+        batch_size, seq_len, dim = tensor.shape
+        tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
+        tensor = tensor.permute(0, 2, 1, 3)
+
+        if out_dim == 3:
+            tensor = tensor.reshape(batch_size * head_size, seq_len, dim // head_size)
+
+        return tensor
+
+    def get_attention_scores(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        attention_mask: torch.Tensor = None,
+    ) -> torch.Tensor:
+        r"""
+        Compute the attention scores.
+
+        Args:
+            query (`torch.Tensor`): The query tensor.
+            key (`torch.Tensor`): The key tensor.
+            attention_mask (`torch.Tensor`, *optional*): The attention mask to use. If `None`, no mask is applied.
+
+        Returns:
+            `torch.Tensor`: The attention probabilities/scores.
+        """
+        dtype = query.dtype
+        if self.upcast_attention:
+            query = query.float()
+            key = key.float()
+
+        if attention_mask is None:
+            baddbmm_input = torch.empty(
+                query.shape[0],
+                query.shape[1],
+                key.shape[1],
+                dtype=query.dtype,
+                device=query.device,
+            )
+            beta = 0
+        else:
+            baddbmm_input = attention_mask
+            beta = 1
+
+        attention_scores = torch.baddbmm(
+            baddbmm_input,
+            query,
+            key.transpose(-1, -2),
+            beta=beta,
+            alpha=self.scale,
+        )
+        del baddbmm_input
+
+        if self.upcast_softmax:
+            attention_scores = attention_scores.float()
+
+        attention_probs = attention_scores.softmax(dim=-1)
+        del attention_scores
+
+        attention_probs = attention_probs.to(dtype)
+
+        return attention_probs
+
+    def prepare_attention_mask(
+        self,
+        attention_mask: torch.Tensor,
+        target_length: int,
+        batch_size: int,
+        out_dim: int = 3,
+    ) -> torch.Tensor:
+        r"""
+        Prepare the attention mask for the attention computation.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                The attention mask to prepare.
+            target_length (`int`):
+                The target length of the attention mask. This is the length of the attention mask after padding.
+            batch_size (`int`):
+                The batch size, which is used to repeat the attention mask.
+            out_dim (`int`, *optional*, defaults to `3`):
+                The output dimension of the attention mask. Can be either `3` or `4`.
+
+        Returns:
+            `torch.Tensor`: The prepared attention mask.
+        """
+        head_size = self.heads
+        if attention_mask is None:
+            return attention_mask
+
+        current_length: int = attention_mask.shape[-1]
+        if current_length != target_length:
+            if attention_mask.device.type == "mps":
+                # HACK: MPS: Does not support padding by greater than dimension of input tensor.
+                # Instead, we can manually construct the padding tensor.
+                padding_shape = (
+                    attention_mask.shape[0],
+                    attention_mask.shape[1],
+                    target_length,
+                )
+                padding = torch.zeros(
+                    padding_shape,
+                    dtype=attention_mask.dtype,
+                    device=attention_mask.device,
+                )
+                attention_mask = torch.cat([attention_mask, padding], dim=2)
+            else:
+                # TODO: for pipelines such as stable-diffusion, padding cross-attn mask:
+                #       we want to instead pad by (0, remaining_length), where remaining_length is:
+                #       remaining_length: int = target_length - current_length
+                # TODO: re-enable tests/models/test_models_unet_2d_condition.py#test_model_xattn_padding
+                attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
+
+        if out_dim == 3:
+            if attention_mask.shape[0] < batch_size * head_size:
+                attention_mask = attention_mask.repeat_interleave(head_size, dim=0)
+        elif out_dim == 4:
+            attention_mask = attention_mask.unsqueeze(1)
+            attention_mask = attention_mask.repeat_interleave(head_size, dim=1)
+
+        return attention_mask
+
+    def norm_encoder_hidden_states(
+        self, encoder_hidden_states: torch.Tensor
+    ) -> torch.Tensor:
+        r"""
+        Normalize the encoder hidden states. Requires `self.norm_cross` to be specified when constructing the
+        `Attention` class.
+
+        Args:
+            encoder_hidden_states (`torch.Tensor`): Hidden states of the encoder.
+
+        Returns:
+            `torch.Tensor`: The normalized encoder hidden states.
+        """
+        assert (
+            self.norm_cross is not None
+        ), "self.norm_cross must be defined to call self.norm_encoder_hidden_states"
+
+        if isinstance(self.norm_cross, nn.LayerNorm):
+            encoder_hidden_states = self.norm_cross(encoder_hidden_states)
+        elif isinstance(self.norm_cross, nn.GroupNorm):
+            # Group norm norms along the channels dimension and expects
+            # input to be in the shape of (N, C, *). In this case, we want
+            # to norm along the hidden dimension, so we need to move
+            # (batch_size, sequence_length, hidden_size) ->
+            # (batch_size, hidden_size, sequence_length)
+            encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
+            encoder_hidden_states = self.norm_cross(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
+        else:
+            assert False
+
+        return encoder_hidden_states
+
+    @torch.no_grad()
+    def fuse_projections(self, fuse=True):
+        is_cross_attention = self.cross_attention_dim != self.query_dim
+        device = self.to_q.weight.data.device
+        dtype = self.to_q.weight.data.dtype
+
+        if not is_cross_attention:
+            # fetch weight matrices.
+            concatenated_weights = torch.cat(
+                [self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data]
+            )
+            in_features = concatenated_weights.shape[1]
+            out_features = concatenated_weights.shape[0]
+
+            # create a new single projection layer and copy over the weights.
+            self.to_qkv = self.linear_cls(
+                in_features, out_features, bias=False, device=device, dtype=dtype
+            )
+            self.to_qkv.weight.copy_(concatenated_weights)
+
+        else:
+            concatenated_weights = torch.cat(
+                [self.to_k.weight.data, self.to_v.weight.data]
+            )
+            in_features = concatenated_weights.shape[1]
+            out_features = concatenated_weights.shape[0]
+
+            self.to_kv = self.linear_cls(
+                in_features, out_features, bias=False, device=device, dtype=dtype
+            )
+            self.to_kv.weight.copy_(concatenated_weights)
+
+        self.fused_projections = fuse
+
+
+class AttnProcessor:
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.Tensor:
+        residual = hidden_states
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(
+                batch_size, channel, height * width
+            ).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape
+            if encoder_hidden_states is None
+            else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(
+            attention_mask, sequence_length, batch_size
+        )
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(
+                1, 2
+            )
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(
+                encoder_hidden_states
+            )
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(
+                batch_size, channel, height, width
+            )
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class AttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        residual = hidden_states
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(
+                batch_size, channel, height * width
+            ).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape
+            if encoder_hidden_states is None
+            else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(
+                attention_mask, sequence_length, batch_size
+            )
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(
+                batch_size, attn.heads, -1, attention_mask.shape[-1]
+            )
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(
+                1, 2
+            )
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(
+                encoder_hidden_states
+            )
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(
+            batch_size, -1, attn.heads * head_dim
+        )
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(
+                batch_size, channel, height, width
+            )
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
@@ -0,0 +1,334 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# --------
+#
+# Modified 2024 by the Tripo AI and Stability AI Team.
+#
+# Copyright (c) 2024 Tripo AI & Stability AI
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from .attention import Attention
+
+
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",
+        final_dropout: bool = False,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        assert norm_type == "layer_norm"
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=(
+                    cross_attention_dim if not double_self_attention else None
+                ),
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+        )
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        norm_hidden_states = self.norm1(hidden_states)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=(
+                encoder_hidden_states if self.only_cross_attention else None
+            ),
+            attention_mask=attention_mask,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = self.norm2(hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
+                raise ValueError(
+                    f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+                )
+
+            num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
+            ff_output = torch.cat(
+                [
+                    self.ff(hid_slice)
+                    for hid_slice in norm_hidden_states.chunk(
+                        num_chunks, dim=self._chunk_dim
+                    )
+                ],
+                dim=self._chunk_dim,
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        hidden_states = ff_output + hidden_states
+
+        return hidden_states
+
+
+class FeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        dim (`int`): The number of channels in the input.
+        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
+        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: int = 4,
+        dropout: float = 0.0,
+        activation_fn: str = "geglu",
+        final_dropout: bool = False,
+    ):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+        linear_cls = nn.Linear
+
+        if activation_fn == "gelu":
+            act_fn = GELU(dim, inner_dim)
+        if activation_fn == "gelu-approximate":
+            act_fn = GELU(dim, inner_dim, approximate="tanh")
+        elif activation_fn == "geglu":
+            act_fn = GEGLU(dim, inner_dim)
+        elif activation_fn == "geglu-approximate":
+            act_fn = ApproximateGELU(dim, inner_dim)
+
+        self.net = nn.ModuleList([])
+        # project in
+        self.net.append(act_fn)
+        # project dropout
+        self.net.append(nn.Dropout(dropout))
+        # project out
+        self.net.append(linear_cls(inner_dim, dim_out))
+        # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
+        if final_dropout:
+            self.net.append(nn.Dropout(dropout))
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        for module in self.net:
+            hidden_states = module(hidden_states)
+        return hidden_states
+
+
+class GELU(nn.Module):
+    r"""
+    GELU activation function with tanh approximation support with `approximate="tanh"`.
+
+    Parameters:
+        dim_in (`int`): The number of channels in the input.
+        dim_out (`int`): The number of channels in the output.
+        approximate (`str`, *optional*, defaults to `"none"`): If `"tanh"`, use tanh approximation.
+    """
+
+    def __init__(self, dim_in: int, dim_out: int, approximate: str = "none"):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out)
+        self.approximate = approximate
+
+    def gelu(self, gate: torch.Tensor) -> torch.Tensor:
+        if gate.device.type != "mps":
+            return F.gelu(gate, approximate=self.approximate)
+        # mps: gelu is not implemented for float16
+        return F.gelu(gate.to(dtype=torch.float32), approximate=self.approximate).to(
+            dtype=gate.dtype
+        )
+
+    def forward(self, hidden_states):
+        hidden_states = self.proj(hidden_states)
+        hidden_states = self.gelu(hidden_states)
+        return hidden_states
+
+
+class GEGLU(nn.Module):
+    r"""
+    A variant of the gated linear unit activation function from https://arxiv.org/abs/2002.05202.
+
+    Parameters:
+        dim_in (`int`): The number of channels in the input.
+        dim_out (`int`): The number of channels in the output.
+    """
+
+    def __init__(self, dim_in: int, dim_out: int):
+        super().__init__()
+        linear_cls = nn.Linear
+
+        self.proj = linear_cls(dim_in, dim_out * 2)
+
+    def gelu(self, gate: torch.Tensor) -> torch.Tensor:
+        if gate.device.type != "mps":
+            return F.gelu(gate)
+        # mps: gelu is not implemented for float16
+        return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
+
+    def forward(self, hidden_states, scale: float = 1.0):
+        args = ()
+        hidden_states, gate = self.proj(hidden_states, *args).chunk(2, dim=-1)
+        return hidden_states * self.gelu(gate)
+
+
+class ApproximateGELU(nn.Module):
+    r"""
+    The approximate form of Gaussian Error Linear Unit (GELU). For more details, see section 2:
+    https://arxiv.org/abs/1606.08415.
+
+    Parameters:
+        dim_in (`int`): The number of channels in the input.
+        dim_out (`int`): The number of channels in the output.
+    """
+
+    def __init__(self, dim_in: int, dim_out: int):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        return x * torch.sigmoid(1.702 * x)
@@ -0,0 +1,219 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# --------
+#
+# Modified 2024 by the Tripo AI and Stability AI Team.
+#
+# Copyright (c) 2024 Tripo AI & Stability AI
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ...utils import BaseModule
+from .basic_transformer_block import BasicTransformerBlock
+
+
+class Transformer1D(BaseModule):
+    @dataclass
+    class Config(BaseModule.Config):
+        num_attention_heads: int = 16
+        attention_head_dim: int = 88
+        in_channels: Optional[int] = None
+        out_channels: Optional[int] = None
+        num_layers: int = 1
+        dropout: float = 0.0
+        norm_num_groups: int = 32
+        cross_attention_dim: Optional[int] = None
+        attention_bias: bool = False
+        activation_fn: str = "geglu"
+        only_cross_attention: bool = False
+        double_self_attention: bool = False
+        upcast_attention: bool = False
+        norm_type: str = "layer_norm"
+        norm_elementwise_affine: bool = True
+        gradient_checkpointing: bool = False
+
+    cfg: Config
+
+    def configure(self) -> None:
+        self.num_attention_heads = self.cfg.num_attention_heads
+        self.attention_head_dim = self.cfg.attention_head_dim
+        inner_dim = self.num_attention_heads * self.attention_head_dim
+
+        linear_cls = nn.Linear
+
+        # 2. Define input layers
+        self.in_channels = self.cfg.in_channels
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=self.cfg.norm_num_groups,
+            num_channels=self.cfg.in_channels,
+            eps=1e-6,
+            affine=True,
+        )
+        self.proj_in = linear_cls(self.cfg.in_channels, inner_dim)
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    self.num_attention_heads,
+                    self.attention_head_dim,
+                    dropout=self.cfg.dropout,
+                    cross_attention_dim=self.cfg.cross_attention_dim,
+                    activation_fn=self.cfg.activation_fn,
+                    attention_bias=self.cfg.attention_bias,
+                    only_cross_attention=self.cfg.only_cross_attention,
+                    double_self_attention=self.cfg.double_self_attention,
+                    upcast_attention=self.cfg.upcast_attention,
+                    norm_type=self.cfg.norm_type,
+                    norm_elementwise_affine=self.cfg.norm_elementwise_affine,
+                )
+                for d in range(self.cfg.num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = (
+            self.cfg.in_channels
+            if self.cfg.out_channels is None
+            else self.cfg.out_channels
+        )
+
+        self.proj_out = linear_cls(inner_dim, self.cfg.in_channels)
+
+        self.gradient_checkpointing = self.cfg.gradient_checkpointing
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+    ):
+        """
+        The [`Transformer1DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+
+        Returns:
+            torch.FloatTensor
+        """
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (
+                1 - encoder_attention_mask.to(hidden_states.dtype)
+            ) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 1. Input
+        batch, _, seq_len = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        inner_dim = hidden_states.shape[1]
+        hidden_states = hidden_states.permute(0, 2, 1).reshape(
+            batch, seq_len, inner_dim
+        )
+        hidden_states = self.proj_in(hidden_states)
+
+        # 2. Blocks
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing:
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    block,
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    use_reentrant=False,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+
+        # 3. Output
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = (
+            hidden_states.reshape(batch, seq_len, inner_dim)
+            .permute(0, 2, 1)
+            .contiguous()
+        )
+
+        output = hidden_states + residual
+
+        return output
@@ -0,0 +1,218 @@
+import math
+import os
+from dataclasses import dataclass, field
+from typing import List, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+import trimesh
+from einops import rearrange
+from huggingface_hub import hf_hub_download
+from omegaconf import OmegaConf
+from PIL import Image
+
+from .models.isosurface import MarchingCubeHelper
+from .utils import (
+    BaseModule,
+    ImagePreprocessor,
+    find_class,
+    get_spherical_cameras,
+    scale_tensor,
+)
+
+
+class TSR(BaseModule):
+    @dataclass
+    class Config(BaseModule.Config):
+        cond_image_size: int
+
+        image_tokenizer_cls: str
+        image_tokenizer: dict
+
+        tokenizer_cls: str
+        tokenizer: dict
+
+        backbone_cls: str
+        backbone: dict
+
+        post_processor_cls: str
+        post_processor: dict
+
+        decoder_cls: str
+        decoder: dict
+
+        renderer_cls: str
+        renderer: dict
+
+    cfg: Config
+
+    @classmethod
+    def from_pretrained(
+        cls, pretrained_model_name_or_path: str, config_name: str, weight_name: str
+    ):
+        if os.path.isdir(pretrained_model_name_or_path):
+            config_path = os.path.join(pretrained_model_name_or_path, config_name)
+            weight_path = os.path.join(pretrained_model_name_or_path, weight_name)
+        else:
+            config_path = hf_hub_download(
+                repo_id=pretrained_model_name_or_path, filename=config_name
+            )
+            weight_path = hf_hub_download(
+                repo_id=pretrained_model_name_or_path, filename=weight_name
+            )
+
+        cfg = OmegaConf.load(config_path)
+        OmegaConf.resolve(cfg)
+        model = cls(cfg)
+        ckpt = torch.load(weight_path, map_location="cpu")
+        model.load_state_dict(ckpt)
+        return model
+
+    @classmethod
+    def from_pretrained_custom(
+            cls, weight_path: str, config_path: str
+    ):
+        cfg = OmegaConf.load(config_path)
+        OmegaConf.resolve(cfg)
+        model = cls(cfg)
+        ckpt = torch.load(weight_path, map_location="cpu")
+        model.load_state_dict(ckpt)
+        return model
+
+    def configure(self):
+        self.image_tokenizer = find_class(self.cfg.image_tokenizer_cls)(
+            self.cfg.image_tokenizer
+        )
+        self.tokenizer = find_class(self.cfg.tokenizer_cls)(self.cfg.tokenizer)
+        self.backbone = find_class(self.cfg.backbone_cls)(self.cfg.backbone)
+        self.post_processor = find_class(self.cfg.post_processor_cls)(
+            self.cfg.post_processor
+        )
+        self.decoder = find_class(self.cfg.decoder_cls)(self.cfg.decoder)
+        self.renderer = find_class(self.cfg.renderer_cls)(self.cfg.renderer)
+        self.image_processor = ImagePreprocessor()
+        self.isosurface_helper = None
+
+    def forward(
+        self,
+        image: Union[
+            PIL.Image.Image,
+            np.ndarray,
+            torch.FloatTensor,
+            List[PIL.Image.Image],
+            List[np.ndarray],
+            List[torch.FloatTensor],
+        ],
+        device: str,
+    ) -> torch.FloatTensor:
+        rgb_cond = self.image_processor(image, self.cfg.cond_image_size)[:, None].to(
+            device
+        )
+        batch_size = rgb_cond.shape[0]
+
+        input_image_tokens: torch.Tensor = self.image_tokenizer(
+            rearrange(rgb_cond, "B Nv H W C -> B Nv C H W", Nv=1),
+        )
+
+        input_image_tokens = rearrange(
+            input_image_tokens, "B Nv C Nt -> B (Nv Nt) C", Nv=1
+        )
+
+        tokens: torch.Tensor = self.tokenizer(batch_size)
+
+        tokens = self.backbone(
+            tokens,
+            encoder_hidden_states=input_image_tokens,
+        )
+
+        scene_codes = self.post_processor(self.tokenizer.detokenize(tokens))
+        return scene_codes
+
+    def render(
+        self,
+        scene_codes,
+        n_views: int,
+        elevation_deg: float = 0.0,
+        camera_distance: float = 1.9,
+        fovy_deg: float = 40.0,
+        height: int = 256,
+        width: int = 256,
+        return_type: str = "pil",
+    ):
+        rays_o, rays_d = get_spherical_cameras(
+            n_views, elevation_deg, camera_distance, fovy_deg, height, width
+        )
+        rays_o, rays_d = rays_o.to(scene_codes.device), rays_d.to(scene_codes.device)
+
+        def process_output(image: torch.FloatTensor):
+            if return_type == "pt":
+                return image
+            elif return_type == "np":
+                return image.detach().cpu().numpy()
+            elif return_type == "pil":
+                return Image.fromarray(
+                    (image.detach().cpu().numpy() * 255.0).astype(np.uint8)
+                )
+            else:
+                raise NotImplementedError
+
+        images = []
+        for scene_code in scene_codes:
+            images_ = []
+            for i in range(n_views):
+                with torch.no_grad():
+                    image = self.renderer(
+                        self.decoder, scene_code, rays_o[i], rays_d[i]
+                    )
+                images_.append(process_output(image))
+            images.append(images_)
+
+        return images
+
+    def set_marching_cubes_resolution(self, resolution: int):
+        if (
+            self.isosurface_helper is not None
+            and self.isosurface_helper.resolution == resolution
+        ):
+            return
+        self.isosurface_helper = MarchingCubeHelper(resolution)
+
+    def extract_mesh(self, scene_codes, resolution: int = 256, threshold: float = 25.0,callback=None):
+        self.set_marching_cubes_resolution(resolution)
+        meshes = []
+        for scene_code in scene_codes:
+            with torch.no_grad():
+                density = self.renderer.query_triplane(
+                    self.decoder,
+                    scale_tensor(
+                        self.isosurface_helper.grid_vertices.to(scene_codes.device),
+                        self.isosurface_helper.points_range,
+                        (-self.renderer.cfg.radius, self.renderer.cfg.radius),
+                    ),
+                    scene_code,
+                )["density_act"]
+            v_pos, t_pos_idx = self.isosurface_helper(-(density - threshold))
+            v_pos = scale_tensor(
+                v_pos,
+                self.isosurface_helper.points_range,
+                (-self.renderer.cfg.radius, self.renderer.cfg.radius),
+            )
+            with torch.no_grad():
+                color = self.renderer.query_triplane(
+                    self.decoder,
+                    v_pos,
+                    scene_code,
+                )["color"]
+            mesh = trimesh.Trimesh(
+                vertices=v_pos.cpu().numpy(),
+                faces=t_pos_idx.cpu().numpy(),
+                vertex_colors=color.cpu().numpy(),
+            )
+            meshes.append(mesh)
+
+            if callback:
+                callback(len(meshes))
+
+        return meshes
@@ -0,0 +1,475 @@
+import importlib
+import math
+from collections import defaultdict
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import imageio
+import numpy as np
+import PIL.Image
+#import rembg
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import trimesh
+from omegaconf import DictConfig, OmegaConf
+#from PIL import Image
+
+
+def parse_structured(fields: Any, cfg: Optional[Union[dict, DictConfig]] = None) -> Any:
+    scfg = OmegaConf.merge(OmegaConf.structured(fields), cfg)
+    return scfg
+
+
+def find_class(cls_string):
+    module_string = ".".join(cls_string.split(".")[:-1])
+    cls_name = cls_string.split(".")[-1]
+    module = importlib.import_module(module_string, package=None)
+    cls = getattr(module, cls_name)
+    return cls
+
+
+def get_intrinsic_from_fov(fov, H, W, bs=-1):
+    focal_length = 0.5 * H / np.tan(0.5 * fov)
+    intrinsic = np.identity(3, dtype=np.float32)
+    intrinsic[0, 0] = focal_length
+    intrinsic[1, 1] = focal_length
+    intrinsic[0, 2] = W / 2.0
+    intrinsic[1, 2] = H / 2.0
+
+    if bs > 0:
+        intrinsic = intrinsic[None].repeat(bs, axis=0)
+
+    return torch.from_numpy(intrinsic)
+
+
+class BaseModule(nn.Module):
+    @dataclass
+    class Config:
+        pass
+
+    cfg: Config  # add this to every subclass of BaseModule to enable static type checking
+
+    def __init__(
+        self, cfg: Optional[Union[dict, DictConfig]] = None, *args, **kwargs
+    ) -> None:
+        super().__init__()
+        self.cfg = parse_structured(self.Config, cfg)
+        self.configure(*args, **kwargs)
+
+    def configure(self, *args, **kwargs) -> None:
+        raise NotImplementedError
+
+
+class ImagePreprocessor:
+    def convert_and_resize(
+        self,
+        image: Union[PIL.Image.Image, np.ndarray, torch.Tensor],
+        size: int,
+    ):
+        if isinstance(image, PIL.Image.Image):
+            image = torch.from_numpy(np.array(image).astype(np.float32) / 255.0)
+        elif isinstance(image, np.ndarray):
+            if image.dtype == np.uint8:
+                image = torch.from_numpy(image.astype(np.float32) / 255.0)
+            else:
+                image = torch.from_numpy(image)
+        elif isinstance(image, torch.Tensor):
+            pass
+
+        batched = image.ndim == 4
+
+        if not batched:
+            image = image[None, ...]
+        image = F.interpolate(
+            image.permute(0, 3, 1, 2),
+            (size, size),
+            mode="bilinear",
+            align_corners=False,
+            antialias=True,
+        ).permute(0, 2, 3, 1)
+        if not batched:
+            image = image[0]
+        return image
+
+    def __call__(
+        self,
+        image: Union[
+            PIL.Image.Image,
+            np.ndarray,
+            torch.FloatTensor,
+            List[PIL.Image.Image],
+            List[np.ndarray],
+            List[torch.FloatTensor],
+        ],
+        size: int,
+    ) -> Any:
+        if isinstance(image, (np.ndarray, torch.FloatTensor)) and image.ndim == 4:
+            image = self.convert_and_resize(image, size)
+        else:
+            if not isinstance(image, list):
+                image = [image]
+            image = [self.convert_and_resize(im, size) for im in image]
+            image = torch.stack(image, dim=0)
+        return image
+
+
+def rays_intersect_bbox(
+    rays_o: torch.Tensor,
+    rays_d: torch.Tensor,
+    radius: float,
+    near: float = 0.0,
+    valid_thresh: float = 0.01,
+):
+    input_shape = rays_o.shape[:-1]
+    rays_o, rays_d = rays_o.view(-1, 3), rays_d.view(-1, 3)
+    rays_d_valid = torch.where(
+        rays_d.abs() < 1e-6, torch.full_like(rays_d, 1e-6), rays_d
+    )
+    if type(radius) in [int, float]:
+        radius = torch.FloatTensor(
+            [[-radius, radius], [-radius, radius], [-radius, radius]]
+        ).to(rays_o.device)
+    radius = (
+        1.0 - 1.0e-3
+    ) * radius  # tighten the radius to make sure the intersection point lies in the bounding box
+    interx0 = (radius[..., 1] - rays_o) / rays_d_valid
+    interx1 = (radius[..., 0] - rays_o) / rays_d_valid
+    t_near = torch.minimum(interx0, interx1).amax(dim=-1).clamp_min(near)
+    t_far = torch.maximum(interx0, interx1).amin(dim=-1)
+
+    # check wheter a ray intersects the bbox or not
+    rays_valid = t_far - t_near > valid_thresh
+
+    t_near[torch.where(~rays_valid)] = 0.0
+    t_far[torch.where(~rays_valid)] = 0.0
+
+    t_near = t_near.view(*input_shape, 1)
+    t_far = t_far.view(*input_shape, 1)
+    rays_valid = rays_valid.view(*input_shape)
+
+    return t_near, t_far, rays_valid
+
+
+def chunk_batch(func: Callable, chunk_size: int, *args, **kwargs) -> Any:
+    if chunk_size <= 0:
+        return func(*args, **kwargs)
+    B = None
+    for arg in list(args) + list(kwargs.values()):
+        if isinstance(arg, torch.Tensor):
+            B = arg.shape[0]
+            break
+    assert (
+        B is not None
+    ), "No tensor found in args or kwargs, cannot determine batch size."
+    out = defaultdict(list)
+    out_type = None
+    # max(1, B) to support B == 0
+    for i in range(0, max(1, B), chunk_size):
+        out_chunk = func(
+            *[
+                arg[i : i + chunk_size] if isinstance(arg, torch.Tensor) else arg
+                for arg in args
+            ],
+            **{
+                k: arg[i : i + chunk_size] if isinstance(arg, torch.Tensor) else arg
+                for k, arg in kwargs.items()
+            },
+        )
+        if out_chunk is None:
+            continue
+        out_type = type(out_chunk)
+        if isinstance(out_chunk, torch.Tensor):
+            out_chunk = {0: out_chunk}
+        elif isinstance(out_chunk, tuple) or isinstance(out_chunk, list):
+            chunk_length = len(out_chunk)
+            out_chunk = {i: chunk for i, chunk in enumerate(out_chunk)}
+        elif isinstance(out_chunk, dict):
+            pass
+        else:
+            print(
+                f"Return value of func must be in type [torch.Tensor, list, tuple, dict], get {type(out_chunk)}."
+            )
+            exit(1)
+        for k, v in out_chunk.items():
+            v = v if torch.is_grad_enabled() else v.detach()
+            out[k].append(v)
+
+    if out_type is None:
+        return None
+
+    out_merged: Dict[Any, Optional[torch.Tensor]] = {}
+    for k, v in out.items():
+        if all([vv is None for vv in v]):
+            # allow None in return value
+            out_merged[k] = None
+        elif all([isinstance(vv, torch.Tensor) for vv in v]):
+            out_merged[k] = torch.cat(v, dim=0)
+        else:
+            raise TypeError(
+                f"Unsupported types in return value of func: {[type(vv) for vv in v if not isinstance(vv, torch.Tensor)]}"
+            )
+
+    if out_type is torch.Tensor:
+        return out_merged[0]
+    elif out_type in [tuple, list]:
+        return out_type([out_merged[i] for i in range(chunk_length)])
+    elif out_type is dict:
+        return out_merged
+
+
+ValidScale = Union[Tuple[float, float], torch.FloatTensor]
+
+
+def scale_tensor(dat: torch.FloatTensor, inp_scale: ValidScale, tgt_scale: ValidScale):
+    if inp_scale is None:
+        inp_scale = (0, 1)
+    if tgt_scale is None:
+        tgt_scale = (0, 1)
+    if isinstance(tgt_scale, torch.FloatTensor):
+        assert dat.shape[-1] == tgt_scale.shape[-1]
+    dat = (dat - inp_scale[0]) / (inp_scale[1] - inp_scale[0])
+    dat = dat * (tgt_scale[1] - tgt_scale[0]) + tgt_scale[0]
+    return dat
+
+
+def get_activation(name) -> Callable:
+    if name is None:
+        return lambda x: x
+    name = name.lower()
+    if name == "none":
+        return lambda x: x
+    elif name == "exp":
+        return lambda x: torch.exp(x)
+    elif name == "sigmoid":
+        return lambda x: torch.sigmoid(x)
+    elif name == "tanh":
+        return lambda x: torch.tanh(x)
+    elif name == "softplus":
+        return lambda x: F.softplus(x)
+    else:
+        try:
+            return getattr(F, name)
+        except AttributeError:
+            raise ValueError(f"Unknown activation function: {name}")
+
+
+def get_ray_directions(
+    H: int,
+    W: int,
+    focal: Union[float, Tuple[float, float]],
+    principal: Optional[Tuple[float, float]] = None,
+    use_pixel_centers: bool = True,
+    normalize: bool = True,
+) -> torch.FloatTensor:
+    """
+    Get ray directions for all pixels in camera coordinate.
+    Reference: https://www.scratchapixel.com/lessons/3d-basic-rendering/
+               ray-tracing-generating-camera-rays/standard-coordinate-systems
+
+    Inputs:
+        H, W, focal, principal, use_pixel_centers: image height, width, focal length, principal point and whether use pixel centers
+    Outputs:
+        directions: (H, W, 3), the direction of the rays in camera coordinate
+    """
+    pixel_center = 0.5 if use_pixel_centers else 0
+
+    if isinstance(focal, float):
+        fx, fy = focal, focal
+        cx, cy = W / 2, H / 2
+    else:
+        fx, fy = focal
+        assert principal is not None
+        cx, cy = principal
+
+    i, j = torch.meshgrid(
+        torch.arange(W, dtype=torch.float32) + pixel_center,
+        torch.arange(H, dtype=torch.float32) + pixel_center,
+        indexing="xy",
+    )
+
+    directions = torch.stack([(i - cx) / fx, -(j - cy) / fy, -torch.ones_like(i)], -1)
+
+    if normalize:
+        directions = F.normalize(directions, dim=-1)
+
+    return directions
+
+
+def get_rays(
+    directions,
+    c2w,
+    keepdim=False,
+    normalize=False,
+) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
+    # Rotate ray directions from camera coordinate to the world coordinate
+    assert directions.shape[-1] == 3
+
+    if directions.ndim == 2:  # (N_rays, 3)
+        if c2w.ndim == 2:  # (4, 4)
+            c2w = c2w[None, :, :]
+        assert c2w.ndim == 3  # (N_rays, 4, 4) or (1, 4, 4)
+        rays_d = (directions[:, None, :] * c2w[:, :3, :3]).sum(-1)  # (N_rays, 3)
+        rays_o = c2w[:, :3, 3].expand(rays_d.shape)
+    elif directions.ndim == 3:  # (H, W, 3)
+        assert c2w.ndim in [2, 3]
+        if c2w.ndim == 2:  # (4, 4)
+            rays_d = (directions[:, :, None, :] * c2w[None, None, :3, :3]).sum(
+                -1
+            )  # (H, W, 3)
+            rays_o = c2w[None, None, :3, 3].expand(rays_d.shape)
+        elif c2w.ndim == 3:  # (B, 4, 4)
+            rays_d = (directions[None, :, :, None, :] * c2w[:, None, None, :3, :3]).sum(
+                -1
+            )  # (B, H, W, 3)
+            rays_o = c2w[:, None, None, :3, 3].expand(rays_d.shape)
+    elif directions.ndim == 4:  # (B, H, W, 3)
+        assert c2w.ndim == 3  # (B, 4, 4)
+        rays_d = (directions[:, :, :, None, :] * c2w[:, None, None, :3, :3]).sum(
+            -1
+        )  # (B, H, W, 3)
+        rays_o = c2w[:, None, None, :3, 3].expand(rays_d.shape)
+
+    if normalize:
+        rays_d = F.normalize(rays_d, dim=-1)
+    if not keepdim:
+        rays_o, rays_d = rays_o.reshape(-1, 3), rays_d.reshape(-1, 3)
+
+    return rays_o, rays_d
+
+
+def get_spherical_cameras(
+    n_views: int,
+    elevation_deg: float,
+    camera_distance: float,
+    fovy_deg: float,
+    height: int,
+    width: int,
+):
+    azimuth_deg = torch.linspace(0, 360.0, n_views + 1)[:n_views]
+    elevation_deg = torch.full_like(azimuth_deg, elevation_deg)
+    camera_distances = torch.full_like(elevation_deg, camera_distance)
+
+    elevation = elevation_deg * math.pi / 180
+    azimuth = azimuth_deg * math.pi / 180
+
+    # convert spherical coordinates to cartesian coordinates
+    # right hand coordinate system, x back, y right, z up
+    # elevation in (-90, 90), azimuth from +x to +y in (-180, 180)
+    camera_positions = torch.stack(
+        [
+            camera_distances * torch.cos(elevation) * torch.cos(azimuth),
+            camera_distances * torch.cos(elevation) * torch.sin(azimuth),
+            camera_distances * torch.sin(elevation),
+        ],
+        dim=-1,
+    )
+
+    # default scene center at origin
+    center = torch.zeros_like(camera_positions)
+    # default camera up direction as +z
+    up = torch.as_tensor([0, 0, 1], dtype=torch.float32)[None, :].repeat(n_views, 1)
+
+    fovy = torch.full_like(elevation_deg, fovy_deg) * math.pi / 180
+
+    lookat = F.normalize(center - camera_positions, dim=-1)
+    right = F.normalize(torch.cross(lookat, up), dim=-1)
+    up = F.normalize(torch.cross(right, lookat), dim=-1)
+    c2w3x4 = torch.cat(
+        [torch.stack([right, up, -lookat], dim=-1), camera_positions[:, :, None]],
+        dim=-1,
+    )
+    c2w = torch.cat([c2w3x4, torch.zeros_like(c2w3x4[:, :1])], dim=1)
+    c2w[:, 3, 3] = 1.0
+
+    # get directions by dividing directions_unit_focal by focal length
+    focal_length = 0.5 * height / torch.tan(0.5 * fovy)
+    directions_unit_focal = get_ray_directions(
+        H=height,
+        W=width,
+        focal=1.0,
+    )
+    directions = directions_unit_focal[None, :, :, :].repeat(n_views, 1, 1, 1)
+    directions[:, :, :, :2] = (
+        directions[:, :, :, :2] / focal_length[:, None, None, None]
+    )
+    # must use normalize=True to normalize directions here
+    rays_o, rays_d = get_rays(directions, c2w, keepdim=True, normalize=True)
+
+    return rays_o, rays_d
+
+
+# def remove_background(
+#     image: PIL.Image.Image,
+#     rembg_session: Any = None,
+#     force: bool = False,
+#     **rembg_kwargs,
+# ) -> PIL.Image.Image:
+#     do_remove = True
+#     if image.mode == "RGBA" and image.getextrema()[3][0] < 255:
+#         do_remove = False
+#     do_remove = do_remove or force
+#     if do_remove:
+#         image = rembg.remove(image, session=rembg_session, **rembg_kwargs)
+#     return image
+
+
+def resize_foreground(
+    image: PIL.Image.Image,
+    ratio: float,
+) -> PIL.Image.Image:
+    image = np.array(image)
+    assert image.shape[-1] == 4
+    alpha = np.where(image[..., 3] > 0)
+    y1, y2, x1, x2 = (
+        alpha[0].min(),
+        alpha[0].max(),
+        alpha[1].min(),
+        alpha[1].max(),
+    )
+    # crop the foreground
+    fg = image[y1:y2, x1:x2]
+    # pad to square
+    size = max(fg.shape[0], fg.shape[1])
+    ph0, pw0 = (size - fg.shape[0]) // 2, (size - fg.shape[1]) // 2
+    ph1, pw1 = size - fg.shape[0] - ph0, size - fg.shape[1] - pw0
+    new_image = np.pad(
+        fg,
+        ((ph0, ph1), (pw0, pw1), (0, 0)),
+        mode="constant",
+        constant_values=((0, 0), (0, 0), (0, 0)),
+    )
+
+    # compute padding according to the ratio
+    new_size = int(new_image.shape[0] / ratio)
+    # pad to size, double side
+    ph0, pw0 = (new_size - size) // 2, (new_size - size) // 2
+    ph1, pw1 = new_size - size - ph0, new_size - size - pw0
+    new_image = np.pad(
+        new_image,
+        ((ph0, ph1), (pw0, pw1), (0, 0)),
+        mode="constant",
+        constant_values=((0, 0), (0, 0), (0, 0)),
+    )
+    new_image = PIL.Image.fromarray(new_image)
+    return new_image
+
+
+def save_video(
+    frames: List[PIL.Image.Image],
+    output_path: str,
+    fps: int = 30,
+):
+    # use imageio to save video
+    frames = [np.array(frame) for frame in frames]
+    writer = imageio.get_writer(output_path, fps=fps)
+    for frame in frames:
+        writer.append_data(frame)
+    writer.close()
+
+
+def to_gradio_3d_orientation(mesh):
+    mesh.apply_transform(trimesh.transformations.rotation_matrix(-np.pi/2, [1, 0, 0]))
+    mesh.apply_scale([1, 1, -1])
+    mesh.apply_transform(trimesh.transformations.rotation_matrix(np.pi/2, [0, 1, 0]))
+    return mesh
@@ -0,0 +1,10 @@
+{
+    "main_pass":
+    [
+        "-n", "-c:v", "libsvtav1",
+        "-pix_fmt", "yuv420p10le",
+        "-crf", "23"
+    ],
+     "extension": "webm",
+     "environment": {"SVT_LOG": "1"}
+}
@@ -0,0 +1,9 @@
+{
+    "main_pass":
+    [
+        "-n", "-c:v", "libx264",
+        "-pix_fmt", "yuv420p",
+        "-crf", "19"
+    ],
+     "extension": "mp4"
+}
@@ -0,0 +1,11 @@
+{
+    "main_pass":
+    [
+        "-n", "-c:v", "libx265",
+        "-pix_fmt", "yuv420p10le",
+        "-preset", "medium",
+        "-crf", "22",
+        "-x265-params", "log-level=quiet"
+    ],
+     "extension": "mp4"
+}
@@ -0,0 +1,9 @@
+{
+    "main_pass":
+    [
+        "-n",
+        "-pix_fmt", "yuv420p",
+        "-crf", "23"
+    ],
+     "extension": "webm"
+}
@@ -0,0 +1,15 @@
+[project]
+name = "comfyui-mixlab-nodes"
+description = "3D, ScreenShareNode & FloatingVideoNode, SpeechRecognition & SpeechSynthesis, GPT, LoadImagesFromLocal, Layers, Other Nodes, ..."
+version = "0.41.0"
+license = "MIT"
+dependencies = ["numpy", "pyOpenSSL", "watchdog", "opencv-python-headless", "matplotlib", "openai", "simple-lama-inpainting", "clip-interrogator==0.6.0", "transformers>=4.36.0", "lark-parser", "imageio-ffmpeg", "rembg[gpu]", "omegaconf==2.3.0", "Pillow>=9.5.0", "einops==0.7.0", "trimesh>=4.0.5", "huggingface-hub", "scikit-image"]
+
+[project.urls]
+Repository = "https://github.com/shadowcz007/comfyui-mixlab-nodes"
+#  Used by Comfy Registry https://comfyregistry.org
+
+[tool.comfy]
+PublisherId = "shadow"
+DisplayName = "comfyui-mixlab-nodes"
+Icon = ""
@@ -3,4 +3,22 @@ pyOpenSSL
 watchdog
 opencv-python-headless
 matplotlib
-openai
+openai
+torchaudio
+# simple-lama-inpainting
+clip-interrogator==0.6.0
+transformers>=4.36.0
+lark-parser
+imageio-ffmpeg
+rembg[gpu]
+omegaconf>=2.3.0
+Pillow>=9.5.0
+einops>=0.7.0
+trimesh>=4.0.5
+huggingface-hub
+scikit-image
+soundfile>=0.12.1
+json-repair
+
+bitsandbytes
+accelerate
@@ -0,0 +1,786 @@
+import { app } from '../../../scripts/app.js'
+import { api } from '../../../scripts/api.js'
+import { $el } from '../../../scripts/ui.js'
+
+import { loadExternalScript, get_position_style } from './common.js'
+
+function setCameraOrbit (modelview, distant, angles, screenNumber) {
+  //2.1 20
+  // const angles = {
+  //   1: -20.0,
+  //   2: -17.9,
+  //   3: -15.8,
+  //   4: -13.7,
+  //   5: -11.6,
+  //   6: -9.5,
+  //   7: -7.4,
+  //   8: -5.3,
+  //   9: -3.2,
+  //   10: -1.1,
+  //   11: 1.1,
+  //   12: 3.2,
+  //   13: 5.3,
+  //   14: 7.4,
+  //   15: 9.5,
+  //   16: 11.6,
+  //   17: 13.7,
+  //   18: 15.8,
+  //   19: 17.9,
+  //   20: 20.0
+  // };
+
+  // 12 3.6
+  // const angles = {
+  //   1: -20.0,
+  //   2: -16.4,
+  //   3: -12.7,
+  //   4: -9.1,
+  //   5: -5.5,
+  //   6: -1.8,
+  //   7: 1.8,
+  //   8: 5.5,
+  //   9: 9.1,
+  //   10: 12.7,
+  //   11: 16.4,
+  //   12: 20.0
+  // }
+
+  const angle = angles[screenNumber]
+
+  let co=modelview.cameraOrbit.split(" ")
+
+  if (angle !== undefined) {
+    
+    modelview.cameraOrbit = `${angle}deg ${co[1]} ${distant}m`
+    console.log(screenNumber, angle)
+  } else {
+    console.error('Invalid screen number')
+  }
+}
+
+const getLocalData = key => {
+  let data = {}
+  try {
+    data = JSON.parse(localStorage.getItem(key)) || {}
+  } catch (error) {
+    return {}
+  }
+  return data
+}
+function getContentTypeFromBase64 (base64Data) {
+  const regex = /^data:(.+);base64,/
+  const matches = base64Data.match(regex)
+  if (matches && matches.length >= 2) {
+    return matches[1]
+  }
+  return null
+}
+function base64ToBlobFromURL (base64URL, contentType) {
+  return fetch(base64URL).then(response => response.blob())
+}
+const setLocalDataOfWin = (key, value) => {
+  localStorage.setItem(key, JSON.stringify(value))
+  // window[key] = value
+}
+
+async function uploadImage_ (blob, fileType = '.svg', filename) {
+  // const blob = await (await fetch(src)).blob();
+  const body = new FormData()
+  body.append(
+    'image',
+    new File([blob], (filename || new Date().getTime()) + fileType)
+  )
+
+  const resp = await api.fetchApi('/upload/image', {
+    method: 'POST',
+    body
+  })
+
+  // console.log(resp)
+  let data = await resp.json()
+  return data
+}
+
+async function uploadImage (blob, fileType = '.svg', filename) {
+  let data = await uploadImage_(blob, fileType, filename)
+  let { name, subfolder } = data
+  let src = api.apiURL(
+    `/view?filename=${encodeURIComponent(
+      name
+    )}&type=input&subfolder=${subfolder}${app.getPreviewFormatParam()}${app.getRandParam()}`
+  )
+  return src
+}
+
+function createImage (url) {
+  let im = new Image()
+  return new Promise((res, rej) => {
+    im.onload = () => res(im)
+    im.src = url
+  })
+}
+
+const parseImage = url => {
+  return new Promise((res, rej) => {
+    fetch(url)
+      .then(response => response.blob())
+      .then(blob => {
+        const reader = new FileReader()
+        reader.onloadend = () => {
+          const base64data = reader.result
+          res(base64data)
+          // 在这里可以将base64数据用于进一步处理或显示图片
+        }
+        reader.readAsDataURL(blob)
+      })
+      .catch(error => {
+        console.log('发生错误:', error)
+      })
+  })
+}
+
+async function extractMaterial (
+  modelViewerVariants,
+  selectMaterial,
+  material_img
+) {
+  // 材质
+  const materialsNames = []
+  for (
+    let index = 0;
+    index < modelViewerVariants.model.materials.length;
+    index++
+  ) {
+    let m = modelViewerVariants.model.materials[index]
+    let thumbUrl
+    try {
+      thumbUrl =
+        await m.pbrMetallicRoughness.baseColorTexture.texture.source.createThumbnail(
+          1024,
+          1024
+        )
+    } catch (error) {}
+    if (thumbUrl)
+      materialsNames.push({
+        value: m.name,
+        text: `#${index} ${m.name}`,
+        index,
+        thumbUrl
+      })
+  }
+
+  selectMaterial.innerHTML = ''
+  material_img.innerHTML = ''
+
+  for (let index = 0; index < materialsNames.length; index++) {
+    const name = materialsNames[index]
+    const option = document.createElement('option')
+    option.value = name.thumbUrl
+    option.textContent = name.text
+    option.setAttribute('data-index', index)
+    selectMaterial.appendChild(option)
+    let img = new Image()
+    img.src = name.thumbUrl
+    // img.setAttribute('data-index',name.index)
+    img.style.width = '40px'
+    material_img.appendChild(img)
+    if (index == 0) {
+      material_img.setAttribute('src', name.thumbUrl)
+    }
+  }
+}
+
+async function changeMaterial (
+  modelViewerVariants,
+  targetMaterial,
+  newImageUrl
+) {
+  const targetTexture = await modelViewerVariants.createTexture(newImageUrl)
+  // 用图片创建纹理
+  targetMaterial.pbrMetallicRoughness.baseColorTexture.setTexture(targetTexture)
+}
+
+function inputFileClick (isFileURL = false, isGlb = false) {
+  return new Promise((res, rej) => {
+    // 创建一个input元素
+    var input = document.createElement('input')
+    input.type = 'file'
+    input.accept = isGlb ? '.glb' : 'image/*'
+
+    // 监听input的change事件
+    input.addEventListener('change', function () {
+      // 获取上传的文件
+      var file = input.files[0]
+
+      if (isFileURL) {
+        res(URL.createObjectURL(file))
+        return
+      }
+
+      // 创建一个FileReader对象来读取文件
+      var reader = new FileReader()
+
+      // 监听FileReader的load事件
+      reader.addEventListener('load', async () => {
+        let base64 = reader.result
+        input.remove()
+        res(base64)
+      })
+
+      // 读取文件
+      reader.readAsDataURL(file)
+    })
+
+    // 触发input的点击事件
+    input.click()
+  })
+}
+
+app.registerExtension({
+  name: 'Mixlab.3D.3DImage',
+  async getCustomWidgets (app) {
+    return {
+      THREED (node, inputName, inputData, app) {
+        // console.log('##node', node, inputName, inputData)
+        const widget = {
+          type: inputData[0], // the type, CHEESE
+          name: inputName, // the name, slice
+          size: [128, 88], // a default size
+          draw (ctx, node, width, y) {},
+          computeSize (...args) {
+            return [128, 88] // a method to compute the current size of the widget
+          },
+          async serializeValue (nodeId, widgetIndex) {
+            let d = getLocalData('_mixlab_3d_image')
+            // console.log('serializeValue', node)
+            if (d && d[node.id]) {
+              let { url, bg, material, images } = d[node.id]
+              let data = {}
+              if (url) {
+                data.image = await parseImage(url)
+              }
+              if (bg) {
+                data.bg_image = await parseImage(bg)
+                if (!data.bg_image.match('data:image/')) {
+                  delete data.bg_image
+                }
+              }
+
+              if (material) {
+                data.material = await parseImage(material)
+              }
+
+              if (images) {
+                data.images = images
+              }
+
+              return JSON.parse(JSON.stringify(data))
+            } else {
+              return {}
+            }
+          }
+        }
+        node.addCustomWidget(widget)
+        return widget
+      }
+    }
+  },
+
+  async init () {
+    await loadExternalScript('/mixlab/app/lib/model-viewer.min.js', 'module')
+  },
+
+  async beforeRegisterNodeDef (nodeType, nodeData, app) {
+    if (nodeType.comfyClass == '3DImage') {
+      console.log('nodeType.comfyClass', nodeType.comfyClass)
+      const orig_nodeCreated = nodeType.prototype.onNodeCreated
+      nodeType.prototype.onNodeCreated = async function () {
+        orig_nodeCreated?.apply(this, arguments)
+        const uploadWidget = this.widgets.filter(w => w.name == 'upload')[0]
+        const widget = {
+          type: 'div',
+          name: 'upload-preview',
+          draw (ctx, node, widget_width, y, widget_height) {
+            Object.assign(
+              this.div.style,
+              get_position_style(ctx, widget_width - 122, 88, node.size[1], 44)
+            )
+          }
+        }
+
+        widget.div = $el('div', {})
+        widget.div.style.width = `120px`
+        document.body.appendChild(widget.div)
+
+        const inputDiv = (key, placeholder, preview) => {
+          let div = document.createElement('div')
+          const ip = document.createElement('button')
+          ip.className = `${'comfy-multiline-input'} ${placeholder}`
+          div.style = `display: flex;
+            align-items: center; 
+            margin: 6px 8px;
+            margin-top: 0;`
+
+          ip.style = `outline: none;
+            border: none;
+            padding: 4px;
+            width: 100px;cursor: pointer;
+            height: 32px;`
+          ip.innerText = placeholder
+          div.appendChild(ip)
+
+          let that = this
+
+          ip.addEventListener('click', async event => {
+            let fileURL = await inputFileClick(true, true)
+
+            // console.log('文件URL: ', fileURL)
+            let html = `<model-viewer  src="${fileURL}" 
+                oncontextmenu="return false;"
+                style="outline:1px solid white"
+                min-field-of-view="0deg" 
+                max-field-of-view="180deg"
+                min-camera-orbit="auto auto 0m"
+                max-camera-orbit="auto auto 1000m"
+                 shadow-intensity="1" 
+                 camera-controls 
+                 touch-action="pan-y">
+                 
+                 <div class="controls">
+                  <div>Variant: <select class="variant"></select></div>
+                  <div>Material: <select class="material"></select></div>
+                  <div>Material: <div class="material_img"> </div></div>
+                  <div>
+                    <button class="bg">BG</button>
+                   
+                  </div>
+                  <div>
+                   <input class="ddcap_distant" type="number" min="1" step="1" value="55">
+                   <input class="total_images" type="number" min="1" max="180" step="1" value="20">
+                   <input class="ddcap_range" type="number" min="0" max="20" step="0.1" value="2.1"> 
+                  <button class="ddcap">Capture Rotational Screenshots</button></div>
+                  
+                  <div><button class="export">Export GLB</button></div>
+                  
+                </div></model-viewer>`
+
+            preview.innerHTML = html
+            if (that.size[1] < 400) {
+              that.setSize([that.size[0], that.size[1] + 300])
+              app.canvas.draw(true, true)
+            }
+
+            const modelViewerVariants = preview.querySelector('model-viewer')
+            const select = preview.querySelector('.variant')
+            const selectMaterial = preview.querySelector('.material')
+            const material_img = preview.querySelector('.material_img')
+            const bg = preview.querySelector('.bg')
+
+            const exportGLB = preview.querySelector('.export')
+
+            const ddcap_distant = preview.querySelector('.ddcap_distant')
+            const total_images = preview.querySelector('.total_images')
+            const ddcap_range = preview.querySelector('.ddcap_range')
+            const ddCap = preview.querySelector('.ddcap')
+            const sleep = (t = 1000) => {
+              return new Promise((res, rej) => {
+                return setTimeout(() => {
+                  res(t)
+                }, t)
+              })
+            }
+
+            async function captureImage (isUrl = true) {
+              let base64Data = modelViewerVariants.toDataURL()
+
+              const contentType = getContentTypeFromBase64(base64Data)
+
+              const blob = await base64ToBlobFromURL(base64Data, contentType)
+
+              if (isUrl) return await uploadImage(blob, '.png')
+              return await uploadImage_(blob, '.png')
+            }
+
+            async function captureImages (
+              ddcap_range = 1,
+              total_images = 12,
+              distant = 0.23
+            ) {
+              //  初始 角度
+              var center = modelViewerVariants.getBoundingBoxCenter().toString()
+              modelViewerVariants.cameraTarget = center
+
+              const startAngle = -((total_images - 1) / 2) * ddcap_range
+              const angles = {}
+
+              for (let i = 0; i < total_images; i++) {
+                angles[i + 1] = startAngle + i * ddcap_range
+              }
+              console.log(angles)
+
+              let frames = []
+
+              modelViewerVariants.removeAttribute('camera-controls')
+
+              for (let i = 0; i < total_images; i++) {
+                setCameraOrbit(modelViewerVariants, distant, angles, i + 1)
+
+                // modelViewerVariants.cameraOrbit = `${currentAngle}deg ${initialCameraOrbit[1]} ${initialCameraOrbit[2]}`
+                await sleep(1000)
+                // console.log(`Capturing image at angle: ${currentAngle}deg`)
+                let file = await captureImage(false)
+                frames.push(file)
+                // currentAngle += angleIncrement
+              }
+              await sleep(1000)
+              // 恢复到初始旋转角度
+              // modelViewerVariants.cameraOrbit = initialCameraOrbit.join(' ')
+              modelViewerVariants.setAttribute('camera-controls', '')
+              return frames
+            }
+
+            ddCap.addEventListener('click', async e => {
+              const distant = Number(ddcap_distant.value), //  23m
+                totalImages = Number(total_images.value),
+                angleIncrement = Number(ddcap_range.value)
+              console.log(angleIncrement, totalImages)
+              let images = await captureImages(
+                angleIncrement,
+                totalImages,
+                distant
+              )
+
+              let dd = getLocalData(key)
+              dd[that.id].images = images
+              setLocalDataOfWin(key, dd)
+            })
+
+            ddcap_distant.addEventListener('input', async e => {
+              // console.log(ddcap_distant.value)
+              const center = modelViewerVariants.getBoundingBoxCenter().toString()
+              modelViewerVariants.cameraTarget = center;
+              const initialCameraOrbit =
+                modelViewerVariants.cameraOrbit.split(' ')
+              modelViewerVariants.cameraOrbit = `${initialCameraOrbit[2]} ${initialCameraOrbit[1]} ${ddcap_distant.value}m`
+              modelViewerVariants.setAttribute('camera-controls', '')
+            })
+
+            // ddcap_range_top.addEventListener('input', async e => {
+            //   // console.log(ddcap_range.value)
+            //   const initialCameraOrbit =
+            //     modelViewerVariants.cameraOrbit.split(' ')
+            //   modelViewerVariants.cameraOrbit = `${initialCameraOrbit[0]} ${ddcap_range_top.value}deg ${initialCameraOrbit[2]}`
+            //   modelViewerVariants.setAttribute('camera-controls', '')
+            // })
+
+            if (modelViewerVariants) {
+              modelViewerVariants.style.width = `${that.size[0] - 48}px`
+              modelViewerVariants.style.height = `${that.size[1] - 48}px`
+            }
+
+            modelViewerVariants.addEventListener('load', async () => {
+              const names = modelViewerVariants.availableVariants
+
+              // 变量
+              for (const name of names) {
+                const option = document.createElement('option')
+                option.value = name
+                option.textContent = name
+                select.appendChild(option)
+              }
+              // Adds a default option.
+              if (names.length === 0) {
+                const option = document.createElement('option')
+                option.value = 'default'
+                option.textContent = 'Default'
+                select.appendChild(option)
+              }
+
+              // 材质
+              extractMaterial(modelViewerVariants, selectMaterial, material_img)
+            })
+
+            let timer = null
+            const delay = 500 // 延迟时间，单位为毫秒
+
+            async function checkCameraChange () {
+              let dd = getLocalData(key)
+
+              //  const fileBlob = new Blob([e.target.result], { type: file.type });
+              let url = await captureImage()
+
+              let bg_blob = await base64ToBlobFromURL(
+                'data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mN88uXrPQAFwwK/6xJ6CQAAAABJRU5ErkJggg=='
+              )
+              let url_bg = await uploadImage(bg_blob, '.png')
+              // console.log('url_bg',url_bg)
+
+              if (!dd[that.id]) {
+                dd[that.id] = { url, bg: url_bg }
+              } else {
+                dd[that.id] = { ...dd[that.id], url }
+              }
+
+              //  材质贴图
+              let thumbUrl = material_img.getAttribute('src')
+              if (thumbUrl) {
+                let tb = await base64ToBlobFromURL(thumbUrl)
+                let tUrl = await uploadImage(tb, '.png')
+                // console.log('材质贴图', tUrl, thumbUrl)
+                dd[that.id].material = tUrl
+              }
+
+              setLocalDataOfWin(key, dd)
+            }
+
+            function startTimer () {
+              if (timer) clearTimeout(timer)
+              timer = setTimeout(checkCameraChange, delay)
+            }
+
+            modelViewerVariants.addEventListener('camera-change', startTimer)
+
+            select.addEventListener('input', async event => {
+              modelViewerVariants.variantName =
+                event.target.value === 'default' ? null : event.target.value
+              // 材质
+              await extractMaterial(
+                modelViewerVariants,
+                selectMaterial,
+                material_img
+              )
+              checkCameraChange()
+            })
+
+            selectMaterial.addEventListener('input', event => {
+              // console.log(selectMaterial.value)
+              material_img.setAttribute('src', selectMaterial.value)
+
+              if (selectMaterial.getAttribute('data-new-material')) {
+                let index =
+                  ~~selectMaterial.selectedOptions[0].getAttribute('data-index')
+                changeMaterial(
+                  modelViewerVariants,
+                  modelViewerVariants.model.materials[index],
+                  selectMaterial.getAttribute('data-new-material')
+                )
+              }
+
+              checkCameraChange()
+            })
+
+            //更新bg
+            const updateBgData = (id, key, url, w, h) => {
+              let dd = getLocalData(key)
+              // console.log(dd[that.id],url)
+              if (!dd[id]) dd[id] = { url: '', bg: url }
+              dd[id] = {
+                ...dd[id],
+                bg: url,
+                bg_w: w,
+                bg_h: h
+              }
+              setLocalDataOfWin(key, dd)
+            }
+
+            bg.addEventListener('click', async () => {
+              //更新bg
+              updateBgData(that.id, key, '', 0, 0)
+              preview.style.backgroundImage = 'none'
+
+              let base64 = await inputFileClick(false, false)
+              // 将读取的文件内容设置为div的背景
+              preview.style.backgroundImage = 'url(' + base64 + ')'
+
+              const contentType = getContentTypeFromBase64(base64)
+
+              const blob = await base64ToBlobFromURL(base64, contentType)
+
+              //  const fileBlob = new Blob([e.target.result], { type: file.type });
+              let bg_url = await uploadImage(blob, '.png')
+              let bg_img = await createImage(base64)
+
+              //更新bg
+              updateBgData(
+                that.id,
+                key,
+                bg_url,
+                bg_img.naturalWidth,
+                bg_img.naturalHeight
+              )
+
+              // 更新尺寸
+              let w = that.size[0] - 128,
+                h = (w * bg_img.naturalHeight) / bg_img.naturalWidth
+
+              if (modelViewerVariants) {
+                modelViewerVariants.style.width = `${w}px`
+                modelViewerVariants.style.height = `${h}px`
+              }
+              preview.style.width = `${w}px`
+            })
+
+            exportGLB.addEventListener('click', async () => {
+              const glTF = await modelViewerVariants.exportScene()
+              const file = new File([glTF], 'export.glb')
+              const link = document.createElement('a')
+              link.download = file.name
+              link.href = URL.createObjectURL(file)
+              link.click()
+            })
+
+            uploadWidget.value = await uploadWidget.serializeValue()
+
+            // 更新尺寸
+            let dd = getLocalData(key)
+            // console.log(dd[that.id],bg_url)
+            if (dd[that.id]) {
+              const { bg_w, bg_h } = dd[that.id]
+              if (bg_h && bg_w) {
+                let w = that.size[0] - 48,
+                  h = (w * bg_h) / bg_w
+
+                if (modelViewerVariants) {
+                  modelViewerVariants.style.width = `${w}px`
+                  modelViewerVariants.style.height = `${h}px`
+                }
+                preview.style.width = `${w}px`
+              }
+            }
+          })
+          return div
+        }
+
+        let preview = document.createElement('div')
+        preview.className = 'preview'
+        preview.style = `margin-top: 12px;
+          display: flex;
+          justify-content: center;
+          align-items: center;background-repeat: no-repeat;
+          background-size: contain;`
+
+        let upload = inputDiv('_mixlab_3d_image', '3D Model', preview)
+
+        widget.div.appendChild(upload)
+        widget.div.appendChild(preview)
+        this.addCustomWidget(widget)
+
+        const onResize = this.onResize
+        let that = this
+        this.onResize = function () {
+          let modelViewerVariants = preview.querySelector('model-viewer')
+
+          // 更新尺寸
+          let dd = getLocalData('_mixlab_3d_image')
+
+          if (dd[that.id]) {
+            const { bg_w, bg_h } = dd[that.id]
+            let w = that.size[0] - 128
+            preview.style.width = `${w}px`
+            console.log('更新尺寸', w)
+
+            if (modelViewerVariants) {
+              modelViewerVariants.style.width = `${w}px`
+              modelViewerVariants.style.height = `${Math.round(
+                that.size[1] * 0.8
+              )}px`
+            }
+
+            if (bg_h && bg_w) {
+              let h = (w * bg_h) / bg_w
+              if (modelViewerVariants) {
+                modelViewerVariants.style.height = `${h}px`
+              }
+            }
+          }
+
+          return onResize?.apply(this, arguments)
+        }
+
+        const onRemoved = this.onRemoved
+        this.onRemoved = () => {
+          upload.remove()
+          preview.remove()
+          widget.div.remove()
+          return onRemoved?.()
+        }
+
+        if (this.onResize) {
+          this.onResize(this.size)
+        }
+        // this.isVirtualNode = true
+        this.serialize_widgets = false //需要保存参数
+      }
+
+      const onExecuted = nodeType.prototype.onExecuted
+      nodeType.prototype.onExecuted = function (message) {
+        const r = onExecuted?.apply?.(this, arguments)
+
+        let div = this.widgets.filter(d => d.div)[0]?.div
+        // console.log('Test', this.widgets)
+
+        let material = message.material[0]
+        if (material) {
+          const { filename, subfolder, type } = material
+          let src = api.apiURL(
+            `/view?filename=${encodeURIComponent(
+              filename
+            )}&type=${type}&subfolder=${subfolder}${app.getPreviewFormatParam()}${app.getRandParam()}`
+          )
+
+          const modelViewerVariants = div.querySelector('model-viewer')
+
+          const selectMaterial = div.querySelector('.material')
+
+          let index =
+            ~~selectMaterial.selectedOptions[0].getAttribute('data-index')
+
+          selectMaterial.setAttribute('data-new-material', src)
+
+          changeMaterial(
+            modelViewerVariants,
+            modelViewerVariants.model.materials[index],
+            src
+          )
+        }
+
+        this.onResize?.(this.size)
+
+        return r
+      }
+    }
+  },
+  async loadedGraphNode (node, app) {
+    // Fires every time a node is constructed
+    // You can modify widgets/add handlers/etc here
+    const sleep = (t = 1000) => {
+      return new Promise((res, rej) => {
+        setTimeout(() => res(1), t)
+      })
+    }
+    if (node.type === '3DImage') {
+      // await sleep(0)
+      let widget = node.widgets.filter(w => w.name === 'upload-preview')[0]
+
+      let dd = getLocalData('_mixlab_3d_image')
+
+      let id = node.id
+      // console.log('3dImage load', node.widgets[0], node.widgets)
+      if (!dd[id]) return
+
+      let { url, bg } = dd[id]
+      if (!url) return
+      // let base64 = await parseImage(url)
+
+      let pre = widget.div.querySelector('.preview')
+      pre.style.width = `${node.size[0] - 24}px`
+      pre.innerHTML = `
+        ${url ? `<img src="${url}" style="width:100%"/>` : ''}
+        `
+      pre.style.backgroundImage = 'url(' + bg + ')'
+
+      const uploadWidget = node.widgets.filter(w => w.name == 'upload')[0]
+      uploadWidget.value = await uploadWidget.serializeValue()
+    }
+  }
+})
@@ -0,0 +1,738 @@
+import { app } from '../../../scripts/app.js'
+import { $el } from '../../../scripts/ui.js'
+import { api } from '../../../scripts/api.js'
+
+import { td_bg } from './td_background.js'
+// console.log('td_bg', td_bg)
+import {
+  getUrl,
+  base64Df,
+  get_position_style,
+  getObjectInfo
+} from './common.js'
+
+//本机安装的插件节点全集
+window._nodesAll = null
+
+const parseImageToBase64 = url => {
+  return new Promise((res, rej) => {
+    fetch(url)
+      .then(response => response.blob())
+      .then(blob => {
+        const reader = new FileReader()
+        reader.onloadend = () => {
+          const base64data = reader.result
+          res(base64data)
+          // 在这里可以将base64数据用于进一步处理或显示图片
+        }
+        reader.readAsDataURL(blob)
+      })
+      .catch(error => {
+        console.log('发生错误:', error)
+      })
+  })
+}
+
+async function drawImageToCanvas (imageUrl, sFactor = 320) {
+  var canvas = document.createElement('canvas')
+  var ctx = canvas.getContext('2d')
+  var img = new Image()
+
+  await new Promise((resolve, reject) => {
+    img.onload = function () {
+      var scaleFactor = sFactor / img.width
+      var canvasWidth = img.width * scaleFactor
+      var canvasHeight = img.height * scaleFactor
+
+      canvas.width = canvasWidth
+      canvas.height = canvasHeight
+
+      ctx.drawImage(img, 0, 0, canvasWidth, canvasHeight)
+
+      resolve()
+    }
+
+    img.onerror = function () {
+      reject(new Error('Failed to load image'))
+    }
+
+    img.src = imageUrl
+  })
+
+  var base64 = canvas.toDataURL('image/jpeg')
+  // console.log(base64); // 输出Base64数据
+  return base64
+  // 可以在这里执行其他操作，比如将Base64数据保存到服务器或显示在页面上
+}
+
+async function extractInputAndOutputData (
+  jsonData,
+  inputIds = [],
+  outputIds = []
+) {
+  // workflow
+  // const workflow=jsonData.workflow;
+  // const nodes=workflow.nodes;
+
+  const data = jsonData.output
+  let input = []
+  let output = []
+  const seed = {}
+  const seedTitle = {}
+
+  for (const id in data) {
+    if (data.hasOwnProperty(id)) {
+      let node = app.graph.getNodeById(id)
+      if (inputIds.includes(id)) {
+        // let node = app.graph.getNodeById(id)
+        let options = {}
+        // 模型
+        try {
+          if (node.type === 'CheckpointLoaderSimple') {
+            options = node.widgets.filter(w => w.name === 'ckpt_name')[0]
+              .options.values
+          } else if (node.type === 'LoraLoader') {
+            options = node.widgets.filter(w => w.name === 'lora_name')[0]
+              .options.values
+          }
+        } catch (error) {}
+
+        if (node.type == 'IntNumber' || node.type == 'FloatSlider') {
+          // min max step
+          let [v, min, max, step] = Array.from(node.widgets, w => w.value)
+          options = { min, max, step }
+          // node.widgets.filter(w => w.type === 'number')[0].options
+        }
+
+        if (node.type == 'PromptSlide') {
+          // min max step
+          options = node.widgets.filter(w => w.type === 'slider')[0].options
+          // 备选的keywords清单
+          try {
+            let keywords = node.widgets.filter(w => w.name === 'upload')[0]
+              .value
+            keywords = JSON.parse(keywords)
+            options.keywords = keywords
+          } catch (error) {
+            console.log(error)
+          }
+        }
+
+        if (node.type == 'ImagesPrompt_') {
+          //图库
+          // console.log('ImagesPrompt_', data[id])
+          let image_base64 = data[id].inputs.image_base64
+          let img_index = 0
+          let imgsData = JSON.parse(data[id].inputs.upload)
+          for (let index = 0; index < imgsData.length; index++) {
+            const imgd = imgsData[index].imgurl
+            imgsData[index].index = index
+            //TODO缩放大小
+            imgsData[index].imgurl = await parseImageToBase64(imgd)
+            if (image_base64 == imgsData[index].imgurl) {
+              img_index = index
+            }
+          }
+          options.images = imgsData
+          delete data[id].inputs.upload
+          delete data[id].inputs.image_base64
+
+          data[id].inputs.imageIndex = img_index
+        }
+
+        if (node.type == 'Color') {
+        }
+
+        // 语音输入的支持
+        if (node.type == 'LoadAndCombinedAudio_') {
+          // if (
+          //   data[id].widgets_values &&
+          //   data[id].widgets_values[0] &&
+          //   data[id].widgets_values[0].base64 &&
+          //   data[id].widgets_values[0].base64.length > 0
+          // ) {
+          //   options.defaultBase64 = data[id].widgets_values[0].base64
+          // }
+
+          input[inputIds.indexOf(id)] = {
+            ...data[id],
+            title: node.title,
+            id,
+            options
+          }
+        }
+
+        if (node.type === 'LoadImage') {
+          // loadImage的mask支持
+          let output = node.outputs.filter(ot => ot.type == 'MASK')[0]
+          if (output.links) {
+            // 有输出
+            options.hasMask = true
+          }
+          // loadImage的默认图，转为base64
+          let imgurl = app.graph.getNodeById(id).imgs[0].src + '&channel=rgb'
+
+          options.defaultImage = await drawImageToCanvas(imgurl, 512)
+          console.log('#loadImage的默认图', options)
+        }
+
+        input[inputIds.indexOf(id)] = {
+          ...data[id],
+          title: node.title,
+          id,
+          options
+        }
+        // input.push()
+      }
+      if (outputIds.includes(id)) {
+        let options = {}
+        //输出的默认图
+        if (
+          node.type === 'SaveImageAndMetadata_' &&
+          app.graph.getNodeById(id).imgs
+        ) {
+          // SaveImageAndMetadata_的默认图，转为base64
+          let imgurl = app.graph.getNodeById(id).imgs[0].src
+
+          options.defaultImage = await drawImageToCanvas(imgurl, 512)
+          console.log('#SaveImageAndMetadata_的默认图', options)
+        }
+
+        // let node = app.graph.getNodeById(id)
+        // output.push()
+        output[outputIds.indexOf(id)] = {
+          ...data[id],
+          title: node.title,
+          id,
+          options
+        }
+      }
+
+      if (
+        node.type === 'KSampler' ||
+        node.type == 'SamplerCustom' ||
+        node.type === 'ChinesePrompt_Mix' ||
+        node.type === 'Seed_' ||
+        node.type === 'SiliconflowLLM' ||
+        node.type === 'ChatGPTOpenAI'
+      ) {
+        // seed 的类型收集
+        try {
+          seed[id] = node.widgets.filter(
+            w => w.name === 'seed' || w.name == 'noise_seed'
+          )[0].linkedWidgets[0].value
+          seedTitle[id] = node.title
+        } catch (error) {}
+      }
+    }
+  }
+
+  // 修复bug，当节点不存在时
+  input = input.filter(i => i)
+  output = output.filter(i => i)
+
+  return { input, output, seed, seedTitle }
+}
+
+async function save_app (json) {
+  let url = getUrl()
+
+  const res = await fetch(`${url}/mixlab/workflow`, {
+    method: 'POST',
+    body: JSON.stringify({
+      data: json,
+      task: 'save_app',
+      filename: json.app.filename,
+      category: json.app.category
+    })
+  })
+  return await res.json()
+}
+
+function downloadJsonFile (jsonData, fileName = 'mix_app.json') {
+  const dataString = JSON.stringify(jsonData)
+  const blob = new Blob([dataString], { type: 'application/json' })
+  const url = URL.createObjectURL(blob)
+
+  const link = document.createElement('a')
+  link.href = url
+  link.download = fileName
+  link.click()
+
+  // 释放URL对象
+  setTimeout(() => {
+    URL.revokeObjectURL(url)
+  }, 0)
+}
+
+async function save (json, download = false, showInfo = true) {
+  let nodesAll = window._nodesAll || (await getObjectInfo())
+
+  console.log('####SAVE', nodesAll, json)
+
+  const name = json[0],
+    version = json[5],
+    share_prefix = json[6], //用于分享的功能扩展
+    link = json[7], //用于创建界面上的跳转链接
+    category = json[8] || '', //用于分类
+    idle_animation = json[9], //用于动画，比如数字人her
+    description = json[4],
+    inputIds = json[2].split('\n').filter(f => f),
+    outputIds = json[3].split('\n').filter(f => f)
+
+  const iconData = json[1][0]
+
+  let { filename, subfolder, type } = iconData
+  let iconUrl = api.apiURL(
+    `/view?filename=${encodeURIComponent(
+      filename
+    )}&type=${type}&subfolder=${subfolder}${app.getPreviewFormatParam()}${app.getRandParam()}`
+  )
+
+  try {
+    let data = await app.graphToPrompt()
+
+    //从output数据里把工作流的节点，插件数据统计出来
+    data.nodesMap = {}
+    for (const id in data.output) {
+      data.nodesMap[data.output[id].class_type] =
+        nodesAll[data.output[id].class_type]
+    }
+
+    let { input, output, seed, seedTitle } = await extractInputAndOutputData(
+      data,
+      inputIds,
+      outputIds
+    )
+
+    let authorAvatar =
+        localStorage.getItem('_mixlab_author_avatar') || base64Df,
+      authorName =
+        localStorage.getItem('_mixlab_author_name') ||
+        localStorage.getItem('Comfy.userName'),
+      authorLink = localStorage.getItem('_mixlab_author_link') || ''
+
+    data.app = {
+      name,
+      description,
+      version,
+      input,
+      output,
+      seed, //控制是fixed 还是random
+      seedTitle,
+      share_prefix,
+      link,
+      category,
+      filename: `${name}_${version}.json`,
+      author: {
+        avatar: authorAvatar,
+        name: authorName,
+        link: authorLink
+      }
+    }
+
+    try {
+      data.app.icon = await drawImageToCanvas(iconUrl)
+    } catch (error) {}
+
+    let images = []
+    if (json[1].length > 1 && idle_animation) {
+      images = Array.from(json[1], j => {
+        let { filename, subfolder, type } = j
+        return api.apiURL(
+          `/view?filename=${encodeURIComponent(
+            filename
+          )}&type=${type}&subfolder=${subfolder}${app.getPreviewFormatParam()}${app.getRandParam()}`
+        )
+      })
+    }
+
+    try {
+      for (let index = 0; index < images.length; index++) {
+        const imgurl = images[index]
+        images[index] = await drawImageToCanvas(imgurl)
+      }
+      data.app.idle_animation = images
+    } catch (error) {}
+
+    // console.log(data.app)
+    // let http_workflow = app.graph.serialize()
+    await save_app(data)
+    if (download) {
+      await downloadJsonFile(data, data.app.filename)
+    }
+
+    if (showInfo) {
+      let open = window.confirm(
+        `You can now access the standalone application on a new page!\n${getUrl()}/mixlab/app${
+          data.app.idle_animation ? '/her.html' : ''
+        }?filename=${encodeURIComponent(
+          data.app.filename
+        )}&category=${encodeURIComponent(data.app.category)}`
+      )
+      if (open)
+        window.open(
+          `${getUrl()}/mixlab/app${
+            data.app.idle_animation ? '/her.html' : ''
+          }?filename=${encodeURIComponent(
+            data.app.filename
+          )}&category=${encodeURIComponent(data.app.category)}`
+        )
+    }
+  } catch (error) {
+    console.log('###error', error)
+  }
+}
+
+function getInputsAndOutputs () {
+  const inputs =
+      `LoadImage LoadImagesToBatch ImagesPrompt_ LoadAndCombinedAudio_ LoadVideoAndSegment_ VHS_LoadVideo CLIPTextEncode PromptSlide TextInput_ Color FloatSlider IntNumber CheckpointLoaderSimple LoraLoader`.split(
+        ' '
+      ),
+    outputs =
+      `SaveTripoSRMesh,PreviewImage,SaveImage,TransparentImage,ShowTextForGPT,CombineAudioVideo,VHS_VideoCombine,VideoCombine_Adv,Image Save,SaveImageAndMetadata_,ClipInterrogator`.split(
+        ','
+      )
+
+  let inputsId = [],
+    outputsId = []
+
+  for (let node of app.graph._nodes) {
+    if (inputs.includes(node.type)) {
+      inputsId.push(node.id)
+    }
+
+    if (outputs.includes(node.type)) {
+      outputsId.push(node.id)
+    }
+  }
+
+  return {
+    input: inputsId,
+    output: outputsId
+  }
+}
+
+app.registerExtension({
+  name: 'Mixlab.utils.AppInfo',
+  init () {
+    if (!window._nodesAll) {
+      getObjectInfo().then(r => (window._nodesAll = r))
+    }
+  },
+  async beforeRegisterNodeDef (nodeType, nodeData, app) {
+    if (nodeType.comfyClass == 'AppInfo') {
+      const orig_nodeCreated = nodeType.prototype.onNodeCreated
+      nodeType.prototype.onNodeCreated = function () {
+        orig_nodeCreated?.apply(this, arguments)
+        // console.log('#orig_nodeCreated', this)
+
+        // 自动计算workflow里哪些节点支持
+        let input_ids = this.widgets.filter(w => w.name == 'input_ids')[0],
+          output_ids = this.widgets.filter(w => w.name == 'output_ids')[0]
+
+        const { input, output } = getInputsAndOutputs()
+        input_ids.value = input.join('\n')
+        output_ids.value = output.join('\n')
+        const widget = {
+          type: 'div',
+          name: 'AppInfoRun',
+          draw (ctx, node, widget_width, y, widget_height) {
+            Object.assign(this.div.style, {
+              ...get_position_style(
+                ctx,
+                widget_width,
+                node.size[1] - widget_height,
+                node.size[1]
+              ),
+              zIndex: 1
+            })
+          }
+        }
+
+        const style = `
+        flex-direction: row;
+        background-color: var(--comfy-input-bg);
+        border-radius: 8px;
+        border-color: var(--border-color);
+        border-style: solid;
+        color: var(--descrip-text);`
+
+        widget.div = $el('div', {})
+
+        const btn = document.createElement('button')
+        btn.innerText = 'Save & Open'
+        btn.style = style
+
+        btn.addEventListener('click', () => {
+          //   console.log('hahhah')
+          if (window._mixlab_app_json) {
+            save(window._mixlab_app_json)
+          } else {
+            alert('Please run the workflow before saving')
+            // app.queuePrompt(0, 1)
+            this.widgets.filter(w => w.name === 'version')[0].value += 1
+          }
+        })
+
+        const download = document.createElement('button')
+        download.innerText = 'Download For App'
+        download.style = style
+        download.style.marginLeft = '12px'
+
+        download.addEventListener('click', () => {
+          //   console.log('hahhah')
+          if (window._mixlab_app_json) {
+            save(window._mixlab_app_json, true)
+          } else {
+            alert('Please run the workflow before saving')
+            // app.queuePrompt(0, 1)
+            this.widgets.filter(w => w.name === 'version')[0].value += 1
+          }
+        })
+
+        //td bg
+        const tdBG = document.createElement('button')
+        tdBG.innerText = 'Canvas Mode'
+        tdBG.style = style
+        tdBG.style.marginLeft = '12px'
+
+        tdBG.addEventListener('click', () => {
+          td_bg.toggle()
+          if (td_bg.running) {
+            tdBG.style.background = 'yellow'
+          } else {
+            tdBG.style.background = 'transparent'
+          }
+        })
+
+        // author
+        let author = document.createElement('div')
+        // author.style=`display: flex`
+
+        let authorAvatar = document.createElement('img')
+        authorAvatar.className = `${'comfy-multiline-input'}`
+        authorAvatar.style = `outline: none;
+          border: none;
+          padding: 4px;
+          width: 32px;
+          cursor: pointer;
+          height: 32px;`
+
+        if (localStorage.getItem('_mixlab_author_avatar')) {
+          authorAvatar.src =
+            localStorage.getItem('_mixlab_author_avatar') || base64Df
+        }
+
+        let authorAvatarUpload = document.createElement('input')
+        authorAvatarUpload.type = 'file'
+        authorAvatarUpload.style = `display:none`
+
+        let authorAvatarInput = document.createElement('div')
+        authorAvatarInput.style = `display: flex;justify-content: flex-start;
+        align-items: center;`
+        let authorAvatarInputLabel = document.createElement('p')
+        authorAvatarInputLabel.innerText = 'Author Avatar'
+        authorAvatarInputLabel.className = `${'comfy-multiline-input'}`
+        authorAvatarInputLabel.style = `font-size:12px`
+
+        authorAvatar.addEventListener('click', e => {
+          authorAvatarUpload.click()
+        })
+
+        authorAvatarInputLabel.addEventListener('click', e => {
+          authorAvatarUpload.click()
+        })
+
+        authorAvatarUpload.addEventListener('change', event => {
+          const file = event.target.files[0]
+          const reader = new FileReader()
+
+          reader.onload = async e => {
+            let im = new Image()
+            im.src = e.target.result
+            authorAvatar.src = e.target.result
+            im.onload = () => {
+              let c = document.createElement('canvas')
+              let ctx = c.getContext('2d')
+              c.width = 72
+              c.height = 72
+              ctx.drawImage(
+                im,
+                0,
+                0,
+                im.naturalWidth,
+                im.naturalHeight,
+                0,
+                0,
+                c.width,
+                c.height
+              )
+              window._mixlab_author_avatar = c.toDataURL()
+              localStorage.setItem(
+                '_mixlab_author_avatar',
+                window._mixlab_author_avatar
+              )
+            }
+          }
+
+          // 以文本形式读取文件
+          reader.readAsDataURL(file)
+        })
+
+        author.appendChild(authorAvatarInput)
+        authorAvatarInput.appendChild(authorAvatarInputLabel)
+        authorAvatarInput.appendChild(authorAvatar)
+        authorAvatarInput.appendChild(authorAvatarUpload)
+
+        let authorName = document.createElement('input')
+        authorName.type = 'text'
+        authorName.value =
+          localStorage.getItem('_mixlab_author_name') ||
+          localStorage.getItem('Comfy.userName')
+        authorName.placeholder = 'author name'
+        authorName.className = `${'comfy-multiline-input'}`
+        authorName.style = `
+          outline: none;
+          border: none;
+          padding: 4px;
+          width: 100%;
+          cursor: pointer;
+          height: 32px;`
+
+        let authorNameInput = document.createElement('div')
+        authorNameInput.style = `display: flex;justify-content: flex-start;
+          align-items: center;`
+        let authorNameInputLabel = document.createElement('p')
+        authorNameInputLabel.innerText = 'Author Name'
+        authorNameInputLabel.className = `${'comfy-multiline-input'}`
+        authorNameInputLabel.style = `font-size:12px;width: 110px`
+
+        authorName.addEventListener('change', e => {
+          window._mixlab_author_name = authorName.value.trim()
+          localStorage.setItem(
+            '_mixlab_author_name',
+            window._mixlab_author_name
+          )
+        })
+
+        author.appendChild(authorNameInput)
+        authorNameInput.appendChild(authorNameInputLabel)
+        authorNameInput.appendChild(authorName)
+
+        // 社交链接
+        let authorLink = document.createElement('input')
+        authorLink.type = 'text'
+        authorLink.value = localStorage.getItem('_mixlab_author_link') || ''
+        authorLink.placeholder = 'author link'
+        authorLink.className = `${'comfy-multiline-input'}`
+        authorLink.style = `
+          outline: none;
+          border: none;
+          padding: 4px;
+          width: 100%;
+          cursor: pointer;
+          height: 32px;`
+
+        let authorLinkInput = document.createElement('div')
+        authorLinkInput.style = `display: flex;justify-content: flex-start;
+          align-items: center;`
+        let authorLinkInputLabel = document.createElement('p')
+        authorLinkInputLabel.innerText = 'Author Link'
+        authorLinkInputLabel.className = `${'comfy-multiline-input'}`
+        authorLinkInputLabel.style = `font-size:12px;width: 110px`
+
+        authorLink.addEventListener('change', e => {
+          window._mixlab_author_link = authorLink.value.trim()
+          localStorage.setItem(
+            '_mixlab_author_link',
+            window._mixlab_author_link
+          )
+        })
+
+        author.appendChild(authorLinkInput)
+        authorLinkInput.appendChild(authorLinkInputLabel)
+        authorLinkInput.appendChild(authorLink)
+
+        widget.div.appendChild(author)
+
+        let btns = document.createElement('div')
+
+        widget.div.appendChild(btns)
+
+        btns.appendChild(btn)
+        btns.appendChild(download)
+        btns.appendChild(tdBG)
+
+        document.body.appendChild(widget.div)
+        this.addCustomWidget(widget)
+
+        const onRemoved = this.onRemoved
+        this.onRemoved = () => {
+          widget.div.remove()
+          return onRemoved?.()
+        }
+
+        this.serialize_widgets = true //需要保存参数
+
+        window._mixlab_app_json = null
+      }
+
+      const onExecuted = nodeType.prototype.onExecuted
+      nodeType.prototype.onExecuted = function (message) {
+        onExecuted?.apply(this, arguments)
+        console.log(message.json)
+        window._mixlab_app_json = message.json
+        try {
+          let a = this.widgets.filter(w => w.name === 'AppInfoRun')[0]
+          if (a) {
+            if (!a.value) a.value = 0
+            a.value += 1
+          }
+
+          const div = this.widgets.filter(w => w.div)[0].div
+          Array.from(div.querySelectorAll('button'), b =>
+            b.innerText != 'Canvas Mode' ? (b.style.background = 'yellow') : ''
+          )
+        } catch (error) {}
+      }
+    }
+  },
+  async loadedGraphNode (node, app) {
+    // console.log('#loadedGraphNode1111')
+    window._mixlab_app_json = null //切换workflow需要清空
+    if (node.type === 'AppInfo') {
+      let auto_save = node.widgets.filter(w => w.name == 'auto_save')[0]
+      if (auto_save) {
+        if (!['enable', 'disable'].includes(auto_save.value)) {
+          auto_save.value = 'enable'
+        }
+      }
+
+      // app.canvas.centerOnNode(node)
+      // app.canvas.setZoom(0.45)
+    }
+  }
+})
+
+api.addEventListener('execution_start', async ({ detail }) => {
+  console.log('#execution_start', detail)
+  window._mixlab_app_json = null
+})
+
+api.addEventListener('executed', async ({ detail }) => {
+  console.log('#executed', detail)
+  // window._mixlab_app_json=null;
+  const { output } = getInputsAndOutputs()
+  if (output.includes(parseInt(detail.node))) {
+    let appinfo = app.graph.findNodesByType('AppInfo')[0]
+    if (appinfo) {
+      let auto_save = appinfo.widgets.filter(w => w.name == 'auto_save')[0]
+      if (auto_save?.value === 'enable') {
+        // 自动保存
+        console.log('auto_save')
+        if (window._mixlab_app_json) save(window._mixlab_app_json, false, false)
+      }
+    }
+  }
+})
@@ -19,7 +19,10 @@ function get_position_style (ctx, widget_width, y, node_height) {
  return {
    transformOrigin: '0 0',
    transform: transform,
-    left: `0`,
+    left:
+    document.querySelector('.comfy-menu').style.display === 'none'
+      ? `60px`
+      : `0`,
    top: `0`,
    cursor: 'pointer',
    position: 'absolute',
@@ -68,9 +71,8 @@ function speakText (text) {
 // speakText('Hello, how are you?');
 //                      #MixCopilot

-const start = (element, id, startBtn) => {
-
-  startBtn.className='loading_mixlab'
+const start = (element, id, startBtn, node) => {
+  startBtn.className = 'loading_mixlab'

  window.recognition = new webkitSpeechRecognition()

@@ -95,15 +97,26 @@ const start = (element, id, startBtn) => {
    localStorage.setItem('_mixlab_speech_recognition', JSON.stringify(data))

    if (timeoutId) clearTimeout(timeoutId)
-   
+
    if (!window.recognition) return

    timeoutId = setTimeout(function () {
      console.log('结果传递：：', result)
-      app.queuePrompt(0, 1)
+
+      // 把数据发送到chatgpt的输入prompt里
+      try {
+        const sendToId = node.widgets.filter(
+          w => w.name === 'Send to ChatGPT #'
+        )[0].value
+        app.graph
+          .getNodeById(sendToId)
+          .widgets.filter(w => w.name === 'prompt')[0].value = result
+      } catch (error) {}
+
+      setTimeout(() => app.queuePrompt(0, 1), 100)
      window.recognition?.stop()
-      window.recognition = null;
-      startBtn.className=''
+      window.recognition = null
+      startBtn.className = ''
      startBtn.innerText = 'START'

      timeoutId = null
@@ -114,7 +127,7 @@ const start = (element, id, startBtn) => {
          !window.recognition &&
          window._mixlab_speech_synthesis_onend
        ) {
-          start(element, id, startBtn)
+          start(element, id, startBtn, node)
          startBtn.innerText = 'STOP'
          if (intervalId) {
            clearInterval(intervalId)
@@ -171,13 +184,21 @@ app.registerExtension({
      nodeType.prototype.onNodeCreated = function () {
        orig_nodeCreated?.apply(this, arguments)

+        const sendTo = ComfyWidgets.INT(
+          this,
+          'Send to ChatGPT #',
+          ['INT', { default: 0 }],
+          app
+        )
+        // console.log('sendTo',sendTo)
+
        const widget = {
          type: 'div',
          name: 'chatgptdiv',
          draw (ctx, node, widget_width, y, widget_height) {
            Object.assign(
              this.div.style,
-              get_position_style(ctx, widget_width, 44, node.size[1])
+              get_position_style(ctx, widget_width, 78, node.size[1])
            )
          }
        }
@@ -189,7 +210,14 @@ app.registerExtension({
        const inputDiv = (key, placeholder) => {
          let div = document.createElement('div')
          const startBtn = document.createElement('button')
+
          const textArea = document.createElement('textarea')
+          textArea.placeholder = 'speak text'
+          // sendTo.type='range';
+          // sendTo.min=0;
+          // sendTo.max=2000;
+          // sendTo.step=1;
+          // sendTo.className='comfy-multiline-input'

          textArea.className = `${'comfy-multiline-input'} ${placeholder}`

@@ -201,13 +229,18 @@ app.registerExtension({
          margin: 0px 8px 6px;`

          startBtn.style = `
-            outline: none;
-            border: none;
-            padding: 4px; `
+          margin-top:48px;
+          background-color: var(--comfy-input-bg);
+          border-radius: 8px;
+          border-color: var(--border-color);
+          border-style: solid;
+          color: var(--descrip-text);
+          `

          startBtn.innerText = 'START'

          div.appendChild(startBtn)
+          // div.appendChild(sendTo);
          div.appendChild(textArea)

          startBtn.addEventListener('click', () => {
@@ -215,13 +248,17 @@ app.registerExtension({
              window.recognition.stop()
              window.recognition = null
              startBtn.innerText = 'START'
-              startBtn.className=''
+              startBtn.className = ''
            } else {
-              start(textArea, this.id, startBtn)
+              start(textArea, this.id, startBtn, this)
              startBtn.innerText = 'STOP'
            }
          })

+          // sendTo.addEventListener('change',()=>{
+          //   console.log(sendTo.value)
+          // })
+
          return div
        }

@@ -239,16 +276,69 @@ app.registerExtension({

        this.serialize_widgets = true //需要保存参数
      }
+
+      // const onGraphConfigured=nodeType.prototype.onGraphConfigured;
+      // nodeType.prototype.onGraphConfigured = function (message) {
+      //   onGraphConfigured?.apply(this, arguments)
+      //   console.log('###SpeechRecognition onGraphConfigured',this,message)
+      // }
+
+      const onExecuted = nodeType.prototype.onExecuted
+      nodeType.prototype.onExecuted = function (message) {
+        onExecuted?.apply(this, arguments)
+        // console.log('this.widgets', this.widgets)
+
+        try {
+          // 是否根据start by 开启
+          let open = message.start_by[0] > 0
+          if (open) {
+            const div = this.widgets.filter(w => w.name == 'chatgptdiv')[0].div
+            const startBtn = div.querySelector('button')
+            let textArea = div.querySelector('textarea')
+            if (open && !window.recognition) {
+              start(textArea, this.id, startBtn, this)
+              startBtn.innerText = 'STOP'
+            } else if (!open && window.recognition) {
+              window.recognition.stop()
+              window.recognition = null
+              startBtn.innerText = 'START'
+              startBtn.className = ''
+            }
+          }
+        } catch (error) {
+          console.log('###SpeechRecognition', error)
+        }
+      }
    }
  },
  async loadedGraphNode (node, app) {
    if (node.type === 'SpeechRecognition') {
      let data = getLocalData('_mixlab_speech_recognition')
-      // console.log('_mixlab_speech_recognition', node.widgets)
+      // console.log('_mixlab_speech_recognition', node )
      let div = node.widgets.filter(f => f.type === 'div')[0]
      if (div && data[node.id]) {
        div.div.querySelector('textarea').value = data[node.id]
      }
+
+      try {
+        let open = node.widgets_values[1] > 0
+        if (open) {
+          const div = node.widgets.filter(w => w.name == 'chatgptdiv')[0].div
+          const startBtn = div.querySelector('button')
+          let textArea = div.querySelector('textarea')
+          if (open && !window.recognition) {
+            start(textArea, node.id, startBtn, node)
+            startBtn.innerText = 'STOP'
+          } else if (!open && window.recognition) {
+            window.recognition.stop()
+            window.recognition = null
+            startBtn.innerText = 'START'
+            startBtn.className = ''
+          }
+        }
+      } catch (error) {
+        console.log('###SpeechRecognition', error)
+      }
    }
  }
 })
@@ -309,3 +399,218 @@ app.registerExtension({
    }
  }
 })
+
+// 上传音频转为base64
+async function uploadAndConvertAudio (file) {
+  if (!file) {
+    alert('Please select a WAV file.')
+    return
+  }
+
+  if (file.type !== 'audio/wav') {
+    alert('Only WAV files are supported.')
+    return
+  }
+
+  try {
+    const base64Audio = await readFileAsDataURL(file)
+    return base64Audio
+  } catch (error) {
+    console.error('Error reading file:', error)
+    alert('Error reading file.')
+  }
+}
+
+function readFileAsDataURL (file) {
+  return new Promise((resolve, reject) => {
+    const reader = new FileReader()
+
+    reader.onload = function (event) {
+      resolve(event.target.result)
+    }
+
+    reader.onerror = function (error) {
+      reject(error)
+    }
+
+    reader.readAsDataURL(file)
+  })
+}
+
+const createInputAudioForBatch = (base64, widget) => {
+  // Create an audio element
+  let audio = document.createElement('audio')
+  audio.src = base64
+  audio.controls = true
+  audio.style = 'width: 120px; display: block'
+
+  // Create a delete button
+  let deleteButton = document.createElement('button')
+  deleteButton.textContent = 'Delete'
+
+  deleteButton.style = `cursor: pointer;
+        font-weight: 300;
+        margin: 2px; 
+        margin-left: 10px;
+        color: var(--descrip-text);
+        background-color: var(--comfy-input-bg);
+        border-radius: 8px;
+        border-color: var(--border-color);
+        border-style: solid;height: 30px;min-width: 122px;
+       `
+
+  // Create a container for the audio and delete button
+  let container = document.createElement('div')
+  container.appendChild(audio)
+  container.appendChild(deleteButton)
+  container.style = `display: flex;margin-top: 12px;`
+
+  // Add event listener for the delete button
+  deleteButton.addEventListener('click', e => {
+    let newValue = []
+    let items = widget.value?.base64 || []
+    for (const v of items) {
+      if (v != base64) newValue.push(v)
+    }
+    widget.value.base64 = newValue
+    container.remove()
+  })
+
+  return container
+}
+
+app.registerExtension({
+  name: 'Mixlab.Comfy.LoadAndCombinedAudio_',
+  async getCustomWidgets (app) {
+    return {
+      AUDIOBASE64 (node, inputName, inputData, app) {
+        // console.log('##node', node)
+        const widget = {
+          value: {
+            base64: []
+          }, // 不能[x,x,x]
+          type: inputData[0], // the type
+          name: inputName, // the name, slice
+          size: [128, 32], // a default size
+          draw (ctx, node, width, y) {},
+          computeSize (...args) {
+            return [128, 122] // a method to compute the current size of the widget
+          }
+          // serializeValue (nodeId, widgetIndex) {
+          //   return widget.value
+          // },
+        }
+        //  widget.something = something;          // maybe adds stuff to it
+        node.addCustomWidget(widget) // adds it to the node
+        return widget // and returns it.
+      }
+    }
+  },
+
+  async beforeRegisterNodeDef (nodeType, nodeData, app) {
+    if (nodeType.comfyClass == 'LoadAndCombinedAudio_') {
+      const orig_nodeCreated = nodeType.prototype.onNodeCreated
+
+      nodeType.prototype.onNodeCreated = function () {
+        orig_nodeCreated?.apply(this, arguments)
+
+        let audiosWidget = this.widgets.filter(w => w.name == 'audios')[0]
+
+        const widget = {
+          type: 'div',
+          name: 'audio_base64',
+          draw (ctx, node, widget_width, y, widget_height) {
+            Object.assign(
+              this.div.style,
+              get_position_style(ctx, widget_width, 44, node.size[1])
+            )
+          },
+          serialize: false
+        }
+
+        widget.div = $el('div', {})
+
+        document.body.appendChild(widget.div)
+
+        let audioPreview = document.createElement('div')
+        let audiosDiv = document.createElement('div') //显示图片
+        audiosDiv.className = 'audios_preview'
+        audiosDiv.style = `width: calc(100% - 14px);
+        display: flex;
+        flex-wrap: wrap;
+        padding: 7px;    justify-content: space-between;
+        align-items: center;`
+
+        const btn = document.createElement('button')
+        btn.innerText = 'Upload Audio'
+
+        btn.style = `cursor: pointer;
+        font-weight: 300;
+        margin: 2px; 
+        color: var(--descrip-text);
+        background-color: var(--comfy-input-bg);
+        border-radius: 8px;
+        border-color: var(--border-color);
+        border-style: solid;height: 30px;min-width: 122px;
+       `
+
+        btn.addEventListener('click', e => {
+          e.preventDefault()
+          let inputAudio = document.createElement('input')
+          inputAudio.type = 'file'
+          inputAudio.accept = "audio/*"
+          inputAudio.style.display = 'none'
+          inputAudio.addEventListener('change', async e => {
+            e.preventDefault()
+            const file = e.target.files[0]
+            let base64 = await uploadAndConvertAudio(file)
+            if (!audiosWidget.value) audiosWidget.value = { base64: [] }
+            audiosWidget.value.base64.push(base64)
+
+            let a = createInputAudioForBatch(base64, audiosWidget)
+            audiosDiv.appendChild(a)
+          })
+
+          inputAudio.click()
+          inputAudio.remove()
+        })
+
+        widget.div.appendChild(audioPreview)
+        audioPreview.appendChild(audiosDiv)
+        audioPreview.appendChild(btn)
+        // audioPreview.appendChild(inputAudio)
+
+        this.addCustomWidget(widget)
+
+        // document.addEventListener('wheel', handleMouseWheel)
+
+        const onRemoved = this.onRemoved
+        this.onRemoved = () => {
+          widget.div.remove()
+          try {
+            // document.removeEventListener('wheel', handleMouseWheel)
+          } catch (error) {
+            console.log(error)
+          }
+
+          return onRemoved?.()
+        }
+
+        this.serialize_widgets = true //需要保存参数
+      }
+    }
+  },
+  async loadedGraphNode (node, app) {
+    if (node.type === 'LoadAndCombinedAudio_') {
+      // await sleep(0)
+      let audiosWidget = node.widgets.filter(w => w.name === 'audios')[0]
+      let audioPreview = node.widgets.filter(w => w.name == 'audio_base64')[0]
+
+      let pre = audioPreview.div.querySelector('.audios_preview')
+      for (const d of audiosWidget.value?.base64 || []) {
+        let im = createInputAudioForBatch(d, audiosWidget)
+        pre.appendChild(im)
+      }
+    }
+  }
+})
@@ -0,0 +1,208 @@
+import { getUrl } from './common.js'
+
+async function* completion (url, messages, controller) {
+  let data = {
+    model: 'gpt-3.5-turbo-16k',
+    messages,
+    temperature: 0.05,
+    stream: true
+  }
+  // if (imageNode) {
+  //   data = { ...data, image_data: [imageNode] }
+  // }
+
+  // let controller = new AbortController()
+
+  let response = await fetch(url, {
+    method: 'POST',
+    body: JSON.stringify(data),
+    headers: {
+      Connection: 'keep-alive',
+      'Content-Type': 'application/json',
+      Accept: 'text/event-stream'
+    },
+    signal: controller.signal
+  })
+
+  const reader = response.body.getReader()
+  const decoder = new TextDecoder()
+
+  let content = ''
+  let leftover = '' // Buffer for partially read lines
+
+  try {
+    let cont = true
+    while (cont) {
+      let result = await reader.read()
+      if (result.done) {
+        break
+      }
+
+      // Add any leftover data to the current chunk of data
+      const text = leftover + decoder.decode(result.value)
+
+      // Check if the last character is a line break
+      const endsWithLineBreak = text.endsWith('\n')
+
+      // Split the text into lines
+      let lines = text.split('\n')
+
+      // If the text doesn't end with a line break, then the last line is incomplete
+      // Store it in leftover to be added to the next chunk of data
+      if (!endsWithLineBreak) {
+        leftover = lines.pop()
+      } else {
+        leftover = '' // Reset leftover if we have a line break at the end
+      }
+
+      // Parse all sse events and add them to result
+      const regex = /^(\S+):\s(.*)$/gm
+      for (const line of lines) {
+        const match = regex.exec(line)
+        if (match) {
+          result[match[1]] = match[2]
+          // since we know this is llama.cpp, let's just decode the json in data
+          if (result.data) {
+            result.data = JSON.parse(result.data)
+            // console.log('#result.data',result.data)
+
+            content += result.data.choices[0].delta?.content || ''
+
+            // yield
+            yield result
+
+            // if we got a stop token from server, we will break here
+            if (result.data.choices[0].finish_reason == 'stop') {
+              if (result.data.generation_settings) {
+                // generation_settings = result.data.generation_settings;
+              }
+              cont = false
+              break
+            }
+          }
+        }
+      }
+    }
+  } catch (e) {
+    console.error('llama error: ', e)
+    throw e
+  } finally {
+    controller.abort()
+  }
+
+  return content
+  // return (await response.json()).content
+}
+export async function completion_ (
+  apiKey,
+  url,
+  model_name,
+  messages,
+  controller,
+  callback
+) {
+  let request = await chatCompletion(
+    apiKey,
+    url,
+    model_name,
+    messages,
+    controller
+  )
+  for await (const chunk of request) {
+    if (callback) callback(chunk)
+  }
+}
+
+export async function* chatCompletion (
+  apiKey,
+  api_url,
+  model_name,
+  messages,
+  controller
+) {
+  const mixlabAPI = `${getUrl()}/chat/completions`
+
+  const requestBody = {
+    messages: messages,
+    stream: true,
+    key: apiKey,
+    model_name: model_name,
+    api_url
+  }
+
+  let response = await fetch(mixlabAPI, {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json'
+      // Authorization: `Bearer ${apiKey}`
+    },
+    body: JSON.stringify(requestBody),
+    mode: 'cors', // This is to ensure the request is made with CORS
+    signal: controller.signal
+  })
+
+  const reader = response.body.getReader()
+  const decoder = new TextDecoder()
+
+  let content = ''
+  let leftover = '' // Buffer for partially read lines
+
+  try {
+    let cont = true
+    while (cont) {
+      let result = await reader.read()
+      if (result.done) {
+        break
+      }
+      const text = leftover + decoder.decode(result.value)
+      // Check if the last character is a line break
+      const endsWithLineBreak = text.endsWith('\n')
+
+      // Split the text into lines
+      let lines = text.split('\n')
+
+      // If the text doesn't end with a line break, then the last line is incomplete
+      // Store it in leftover to be added to the next chunk of data
+      if (!endsWithLineBreak) {
+        leftover = lines.pop()
+      } else {
+        leftover = '' // Reset leftover if we have a line break at the end
+      }
+
+      // Parse all sse events and add them to result
+      const regex = /^(\S+):\s(.*)$/gm
+      for (const line of lines) {
+        const match = regex.exec(line)
+        if (match) {
+          result[match[1]] = match[2]
+          // since we know this is llama.cpp, let's just decode the json in data
+          if (result.data) {
+            result.data = JSON.parse(result.data)
+            
+
+            content += result.data.choices[0].delta?.content || ''
+            // console.log('#result.content',content)
+            // yield
+            yield result
+
+            // if we got a stop token from server, we will break here
+            if (result.data.choices[0].finish_reason == 'stop') {
+              if (result.data.generation_settings) {
+                // generation_settings = result.data.generation_settings;
+              }
+              cont = false
+              break
+            }
+          }
+        }
+      }
+    }
+  } catch (e) {
+    console.error('chat error: ', e)
+    throw e
+  } finally {
+    controller.abort()
+  }
+
+  return content
+}
@@ -3,7 +3,7 @@ import { app } from '../../../scripts/app.js'
 const repoOwner = 'shadowcz007' // 替换为仓库的所有者
 const repoName = 'comfyui-mixlab-nodes' // 替换为仓库的名称

-const version = 'v0.3.0'
+const version = 'v0.41.0'

 fetch(`https://api.github.com/repos/${repoOwner}/${repoName}/releases/latest`)
  .then(response => response.json())
@@ -17,7 +17,13 @@ fetch(`https://api.github.com/repos/${repoOwner}/${repoName}/releases/latest`)
      return
    if (latestVersion && latestVersion != version) {
      localStorage.setItem('_mixlab_nodes_vesion', latestVersion)
-      app.ui.dialog.show(`<h4 style="font-size: 18px;">${repoName} <br>
+      app.ui.dialog.show(`<a style="color: white;
+      font-size: 18px;
+      font-weight: 800;
+      letter-spacing: 2px;
+    }"
+    href="https://discord.gg/cXs9vZSqeK">Welcome to Mixlab nodes discord</a>
+    <h4 style="font-size: 18px;">${repoName} <br>
      Latest release version: ${latestVersion}</h4>
      <p>Please proceed to the official repository to download the latest version.</p>
      <a style="color: #2196F3;
@@ -0,0 +1,119 @@
+import { app } from '../../../scripts/app.js'
+// import { api } from '../../../scripts/api.js'
+import { ComfyWidgets } from '../../../scripts/widgets.js'
+import { $el } from '../../../scripts/ui.js'
+
+function getRandomElements (arr, num) {
+  var result = []
+  var len = arr.length
+
+  for (var i = 0; i < num; i++) {
+    var randomIndex = Math.floor(Math.random() * len)
+    result.push(arr[randomIndex])
+  }
+
+  return result
+}
+
+const createPrompt = (node, prompts, items, sample) => {
+  const w = ComfyWidgets['STRING'](
+    node,
+    'text',
+    ['STRING', { multiline: true }],
+    app
+  ).widget
+  w.inputEl.readOnly = true
+  w.inputEl.style.opacity = 0.6
+
+  w.value = typeof prompts === 'string' ? prompts : prompts.join('\n\n')
+
+  const w2 = ComfyWidgets['STRING'](
+    node,
+    'text',
+    ['STRING', { multiline: true }],
+    app
+  ).widget
+  w2.inputEl.readOnly = true
+  w2.inputEl.style.opacity = 0.6
+
+  w2.value = typeof items === 'string' ? items : JSON.stringify(items, null, 2)
+
+  const w3 = ComfyWidgets['STRING'](
+    node,
+    'text',
+    ['STRING', { multiline: true }],
+    app
+  ).widget
+  w3.inputEl.readOnly = true
+  w3.inputEl.style.opacity = 0.6
+  w3.value = typeof sample === 'string' ? sample : sample.join('\n\n')
+}
+
+app.registerExtension({
+  name: 'Mixlab.prompt.ClipInterrogator',
+  async beforeRegisterNodeDef (nodeType, nodeData, app) {
+    if (nodeData.name === 'ClipInterrogator') {
+      function populate (prompts, items, random_samples) {
+        if (this.widgets) {
+          for (let i = 0; i < this.widgets.length; i++) {
+            if (this.widgets[i].type !== 'combo') this.widgets[i].onRemove?.()
+          }
+          this.widgets.length = 2
+        }
+
+        createPrompt(this, prompts, items, random_samples)
+
+        // console.log('ClipInterrogator', w, w2)
+        requestAnimationFrame(() => {
+          const sz = this.computeSize()
+          if (sz[0] < this.size[0]) {
+            sz[0] = this.size[0]
+          }
+          if (sz[1] < this.size[1]) {
+            sz[1] = this.size[1]
+          }
+          this.onResize?.(sz)
+          app.graph.setDirtyCanvas(true, false)
+        })
+      }
+
+      // When the node is executed we will be sent the input text, display this in the widget
+      const onExecuted = nodeType.prototype.onExecuted
+      nodeType.prototype.onExecuted = function (message) {
+        onExecuted?.apply(this, arguments)
+        // console.log('##', message)
+        populate.call(
+          this,
+          message.prompt,
+          message.analysis,
+          message.random_samples
+        )
+      }
+
+      this.serialize_widgets = true //需要保存参数
+    }
+  },
+  async loadedGraphNode (node, app) {
+    // Fires every time a node is constructed
+    // You can modify widgets/add handlers/etc here
+
+    if (node.type === 'ClipInterrogator') {
+      try {
+        
+        let widgets_values = node.widgets_values
+        console.log(widgets_values )
+        try {
+          if (widgets_values[2] && widgets_values[3] && widgets_values[4])
+            createPrompt(
+              node,
+              widgets_values[2],
+              widgets_values[3],
+              widgets_values[4]
+            )
+        } catch (error) {
+          console.log(error)
+        }
+      } catch (error) {}
+    }
+  }
+})
@@ -0,0 +1,227 @@
+export const base64Df =
+  'data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAwAAAAMCAYAAABWdVznAAAAAXNSR0IArs4c6QAAALZJREFUKFOFkLERwjAQBPdbgBkInECGaMLUQDsE0AkRVRAYWqAByxldPPOWHwnw4OBGye1p50UDSoA+W2ABLPN7i+C5dyC6R/uiAUXRQCs0bXoNIu4QPQzAxDKxHoALOrZcqtiyR/T6CXw7+3IGHhkYcy6BOR2izwT8LptG8rbMiCRAUb+CQ6WzQVb0SNOi5Z2/nX35DRyb/ENazhpWKoGwrpD6nICp5c2qogc4of+c7QcrhgF4Aa/aoAFHiL+RAAAAAElFTkSuQmCC'
+
+export function getUrl () {
+  let api_host = `${window.location.hostname}:${window.location.port}`
+  let api_base = ''
+  let url = `${window.location.protocol}//${api_host}${api_base}`
+  return url
+}
+
+// 获得插件/节点的索引数据
+export async function get_nodes_map () {
+  let url = getUrl()
+
+  const res = await fetch(`${url}/mixlab/nodes_map`, {
+    method: 'POST',
+    body: JSON.stringify({
+      data: 'json'
+    })
+  })
+  return await res.json()
+}
+
+// 更新或者获取key
+export const updateLLMAPIKey = async key => {
+  try {
+    const res = await fetch(`${getUrl()}/mixlab/llm_api_key`, {
+      method: 'POST',
+      headers: {
+        'Content-Type': 'application/json'
+      },
+      body: JSON.stringify({
+        key: key || null
+      })
+    })
+
+    const data = await res.json()
+
+    if (!res.ok) {
+      console.error('Error:', data.error)
+      return
+    }
+
+    if (key) {
+      console.log('API key saved successfully:', data.message)
+      return key
+    } else {
+      console.log('Retrieved API key:', data.key)
+      return data.key
+    }
+  } catch (error) {
+    console.error('Request failed:', error)
+  }
+}
+
+//获取当前系统的插件，节点清单
+export function getObjectInfo () {
+  return new Promise(async (resolve, reject) => {
+    let url = getUrl()
+
+    try {
+      const response = await fetch(`${url}/object_info`)
+      const data = await response.json()
+      resolve(data)
+    } catch (error) {
+      reject(error)
+    }
+  })
+}
+
+export function get_position_style (
+  ctx,
+  widget_width,
+  y,
+  node_height,
+  left = 44
+) {
+  const MARGIN = 0 // the margin around the html element
+
+  /* Create a transform that deals with all the scrolling and zooming */
+  const elRect = ctx.canvas.getBoundingClientRect()
+
+  const scaleX = elRect.width / ctx.canvas.width
+  const scaleY = elRect.height / ctx.canvas.height
+
+  const transform = new DOMMatrix()
+    .scaleSelf(scaleX, scaleY)
+    .multiplySelf(ctx.getTransform())
+    .translateSelf(MARGIN, MARGIN + y)
+
+  return {
+    transformOrigin: '0 0',
+    transform: transform,
+    left:
+      document.querySelector('.comfy-menu').style.display === 'none'
+        ? `${left}px`
+        : `0`,
+    top: `0`,
+    cursor: 'pointer',
+    position: 'absolute',
+    maxWidth: `${widget_width - MARGIN * 2}px`,
+    // maxHeight: `${node_height - MARGIN * 2}px`, // we're assuming we have the whole height of the node
+    width: `${widget_width - MARGIN * 2}px`,
+    height: `${node_height * 0.3 - MARGIN * 2}px`,
+    // background: '#EEEEEE',
+    display: 'flex',
+    flexDirection: 'column',
+    // alignItems: 'center',
+    justifyContent: 'flex-start',
+    zIndex: 99
+  }
+}
+
+export function loadCSS (url) {
+  var link = document.createElement('link')
+  link.rel = 'stylesheet'
+  link.type = 'text/css'
+  link.href = url
+  document.getElementsByTagName('head')[0].appendChild(link)
+}
+
+
+export function injectCSS (css) {
+  // 检查页面中是否已经存在具有相同内容的style标签
+  const existingStyle = document.querySelector('style')
+  if (existingStyle && existingStyle.textContent === css) {
+    return // 如果已经存在相同的样式，则不进行注入
+  }
+
+  // 创建一个新的style标签，并将CSS内容注入其中
+  const style = document.createElement('style')
+  style.textContent = css
+
+  // 将style标签插入到页面的head元素中
+  const head = document.querySelector('head')
+  head.appendChild(style)
+}
+
+
+export function loadExternalScript (url, type) {
+  return new Promise((resolve, reject) => {
+    const existingScript = document.querySelector(`script[src="${url}"]`)
+    if (existingScript) {
+      existingScript.onload = () => {
+        resolve()
+      }
+      existingScript.onerror = reject
+      return
+    }
+
+    const script = document.createElement('script')
+    script.src = url
+    if (type) script.type = type // Add this line to load the script as an ES module
+    script.onload = () => {
+      resolve()
+    }
+    script.onerror = reject
+    document.head.appendChild(script)
+  })
+}
+
+export async function getQueue () {
+  try {
+    const res = await fetch(`${getUrl()}/queue`)
+    const data = await res.json()
+    // console.log(data.queue_running,data.queue_pending)
+    return {
+      // Running action uses a different endpoint for cancelling
+      Running: data.queue_running.length,
+      Pending: data.queue_pending.length
+    }
+  } catch (error) {
+    console.error(error)
+    return { Running: 0, Pending: 0 }
+  }
+}
+
+export async function interrupt () {
+  const resp = await fetch(`${getUrl()}/interrupt`, {
+    method: 'POST'
+  })
+}
+
+export async function sleep (t = 200) {
+  return new Promise((res, rej) => {
+    setTimeout(() => {
+      res(true)
+    }, t)
+  })
+}
+
+export function createImage (url) {
+  let im = new Image()
+  return new Promise((res, rej) => {
+    im.onload = () => res(im)
+    im.src = url
+  })
+}
+
+export function convertImageUrlToBase64 (imageUrl) {
+  return fetch(imageUrl)
+    .then(response => response.blob())
+    .then(blob => {
+      return new Promise((resolve, reject) => {
+        const reader = new FileReader()
+        reader.onloadend = () => resolve(reader.result)
+        reader.onerror = reject
+        reader.readAsDataURL(blob)
+      })
+    })
+}
+
+export const getLocalData = key => {
+  let data = {}
+  try {
+    data = JSON.parse(localStorage.getItem(key)) || {}
+  } catch (error) {
+    return {}
+  }
+  return data
+}
+
+export const saveLocalData = (key, id, val) => {
+  let data = getLocalData(key)
+  data[id] = val
+  localStorage.setItem(key, JSON.stringify(data))
+}
@@ -1,277 +1,89 @@
 import { app } from '../../../scripts/app.js'
-// import { api } from '../../../scripts/api.js'
 import { ComfyWidgets } from '../../../scripts/widgets.js'
-import { $el } from '../../../scripts/ui.js'
-
-async function getConfig () {
-  let api_host = `${window.location.hostname}:${window.location.port}`
-  let api_base = ''
-  let url = `${window.location.protocol}//${api_host}${api_base}`
-
-  const res = await fetch(`${url}/mixlab`, {
-    method: 'POST'
-  })
-  return await res.json()
-}
-
-function get_position_style (ctx, widget_width, y, node_height) {
-  const MARGIN = 4 // the margin around the html element
-
-  /* Create a transform that deals with all the scrolling and zooming */
-  const elRect = ctx.canvas.getBoundingClientRect()
-  const transform = new DOMMatrix()
-    .scaleSelf(
-      elRect.width / ctx.canvas.width,
-      elRect.height / ctx.canvas.height
-    )
-    .multiplySelf(ctx.getTransform())
-    .translateSelf(MARGIN, MARGIN + y)
-
-  return {
-    transformOrigin: '0 0',
-    transform: transform,
-    left: `0`,
-    top: `0`,
-    cursor: 'pointer',
-    position: 'absolute',
-    maxWidth: `${widget_width - MARGIN * 2}px`,
-    // maxHeight: `${node_height - MARGIN * 2}px`, // we're assuming we have the whole height of the node
-    width: `${widget_width - MARGIN * 2}px`,
-    // height: `${node_height * 0.3 - MARGIN * 2}px`,
-    // background: '#EEEEEE',
-    display: 'flex',
-    flexDirection: 'column',
-    // alignItems: 'center',
-    justifyContent: 'space-around'
-  }
-}
-
-const getLocalData = key => {
-  let data = {}
-  try {
-    data = JSON.parse(localStorage.getItem(key)) || {}
-  } catch (error) {
-    return {}
-  }
-  return data
-}
-
-app.registerExtension({
-  name: 'Mixlab.GPT.ChatGPTOpenAI',
-  async getCustomWidgets (app) {
-    return {
-      KEY (node, inputName, inputData, app) {
-        // console.log('##node', node)
-        const widget = {
-          type: inputData[0], // the type, CHEESE
-          name: inputName, // the name, slice
-          size: [128, 32], // a default size
-          draw (ctx, node, width, y) {},
-          computeSize (...args) {
-            return [128,32] // a method to compute the current size of the widget
-          },
-          async serializeValue (nodeId, widgetIndex) {
-            let data = getLocalData('_mixlab_api_key')
-            return data[node.id] || 'by Mixlab'
-          }
-        }
-        //  widget.something = something;          // maybe adds stuff to it
-        node.addCustomWidget(widget) // adds it to the node
-        return widget // and returns it.
-      },
-      URL (node, inputName, inputData, app) {
-        // console.log('node', inputName, inputData[0])
-        const widget = {
-          type: inputData[0], // the type, CHEESE
-          name: inputName, // the name, slice
-          size: [128, 32], // a default size
-          draw (ctx, node, width, y) {
-            // a method to draw the widget (ctx is a CanvasRenderingContext2D)
-          },
-          computeSize (...args) {
-            return [128, 32] // a method to compute the current size of the widget
-          },
-          async serializeValue (nodeId, widgetIndex) {
-            let data = getLocalData('_mixlab_api_url')
-            return data[node.id] || 'https://api.openai.com/v1'
-          }
-        }
-        //  widget.something = something;          // maybe adds stuff to it
-        node.addCustomWidget(widget) // adds it to the node
-        return widget // and returns it.
-      }
-    }
-  },
-
-  async beforeRegisterNodeDef (nodeType, nodeData, app) {
-    if (nodeType.comfyClass == 'ChatGPTOpenAI') {
-      const orig_nodeCreated = nodeType.prototype.onNodeCreated
-      nodeType.prototype.onNodeCreated = function () {
-        orig_nodeCreated?.apply(this, arguments)
-
-        const api_key = this.widgets.filter(w => w.name == 'api_key')[0]
-        const api_url = this.widgets.filter(w => w.name == 'api_url')[0]
-
-        console.log('ChatGPTOpenAI nodeData', this.widgets)
-
-        const widget = {
-          type: 'div',
-          name: 'chatgptdiv',
-          draw (ctx, node, widget_width, y, widget_height) {
-            Object.assign(
-              this.div.style,
-              get_position_style(ctx, widget_width, api_key.y, node.size[1])
-            )
-          }
-        }
-
-        widget.div = $el('div', {})
-
-        document.body.appendChild(widget.div)
-
-        const inputDiv = (key, placeholder) => {
-          let div = document.createElement('div')
-          const ip = document.createElement('input')
-          ip.type = placeholder === 'Key' ? 'password' : 'text'
-          ip.className = `${'comfy-multiline-input'} ${placeholder}`
-          div.style = `display: flex;
-          align-items: center; 
-          margin: 6px 8px;
-          margin-top: 0;`
-          ip.placeholder = placeholder
-          ip.value = placeholder
-
-          ip.style = `margin-left: 24px;
-          outline: none;
-          border: none;
-          padding: 4px;width: 100%;`
-          const label = document.createElement('label')
-          label.style = 'font-size: 10px;min-width:32px'
-          label.innerText = placeholder
-          div.appendChild(label)
-          div.appendChild(ip)
-
-          ip.addEventListener('change', () => {
-            let data = getLocalData(key)
-            data[this.id] = ip.value.trim()
-            localStorage.setItem(key, JSON.stringify(data))
-            console.log(this.id, key)
-          })
-          return div
-        }
-
-        let inputKey = inputDiv('_mixlab_api_key', 'Key')
-        let inputUrl = inputDiv('_mixlab_api_url', 'URL')
-
-        widget.div.appendChild(inputKey)
-        widget.div.appendChild(inputUrl)
-
-        this.addCustomWidget(widget)
-
-        const onRemoved = this.onRemoved
-        this.onRemoved = () => {
-          inputUrl.remove()
-          inputKey.remove()
-          widget.div.remove()
-          return onRemoved?.()
-        }
-
-        this.serialize_widgets = true //需要保存参数
-      }
-    }
-  },
-  async loadedGraphNode (node, app) {
-    // Fires every time a node is constructed
-    // You can modify widgets/add handlers/etc here
-
-    if (node.type === 'ChatGPTOpenAI') {
-      let widget = node.widgets.filter(w => w.div)[0]
-
-      let apiKey = getLocalData('_mixlab_api_key'),
-        url = getLocalData('_mixlab_api_url')
-
-      let id = node.id
-
-      // console.log('ChatGPTOpenAI serialize_widgets', this)
-
-      widget.div.querySelector('.Key').value = apiKey[id] || 'by Mixlab'
-      widget.div.querySelector('.URL').value =
-        url[id] || 'https://api.openai.com/v1'
-    }
-  }
-})

 app.registerExtension({
  name: 'Mixlab.GPT.ShowTextForGPT',
-  async beforeRegisterNodeDef(nodeType, nodeData, app) {
-		if (nodeData.name === "ShowTextForGPT") {
-			function populate(text) {
-				if (this.widgets) {
-          
-					const pos = this.widgets.findIndex((w) => w.name === "text");
-					if (pos !== -1) {
-						for (let i = pos; i < this.widgets.length; i++) {
-							this.widgets[i].onRemove?.();
-						}
-						this.widgets.length = pos;
-					}
-				}
-        // console.log('ShowTextForGPT',text)
-				for (let list of text) {
-					const w = ComfyWidgets["STRING"](this, "text", ["STRING", { multiline: true }], app).widget;
-					w.inputEl.readOnly = true;
-					w.inputEl.style.opacity = 0.6;
+  async beforeRegisterNodeDef (nodeType, nodeData, app) {
+    if (nodeData.name === 'ShowTextForGPT') {
+      function populate (text) {
+        text = text.filter(t => t && t?.trim())

-          try {
-            let data=JSON.parse(list);
-            data=Array.from(data,d=>{
-              return {
-                ...d,
-                content:decodeURIComponent(d.content)
-              }
-            })
-            list=JSON.stringify(data,null,2)
-          } catch (error) {
-            // console.log(error)
+        if (this.widgets) {
+          // console.log('#ShowTextForGPT',this.widgets)
+          // const pos = this.widgets.findIndex(w => w.name === 'text')
+          for (let i = 0; i < this.widgets.length; i++) {
+            if (this.widgets[i].name == 'show_text')
+              this.widgets[i].onRemove?.()
+       
          }
+          this.widgets.length = 2
+        }

-					w.value =list;
-        
-				}
+        for (let list of text) {
+          if (list) {
+            // console.log('#####', list)
+            const w = ComfyWidgets['STRING'](
+              this,
+              'show_text',
+              ['STRING', { multiline: true }],
+              app
+            ).widget
+            w.inputEl.readOnly = true
+            w.inputEl.style.opacity = 0.6
+
+            // w.inputEl.style.display='none'
+
+            try {
+              if (typeof list != 'string') {
+                let data = JSON.parse(list)
+                data = Array.from(data, d => {
+                  return {
+                    ...d,
+                    content: decodeURIComponent(d.content)
+                  }
+                })
+                list = JSON.stringify(data, null, 2)
+              }
+            } catch (error) {
+              console.log(error)
+            }
+
+            w.value = list
+          }
+        }
        // console.log('ShowTextForGPT',this.widgets.length)
-				requestAnimationFrame(() => {
-					const sz = this.computeSize();
-					if (sz[0] < this.size[0]) {
-						sz[0] = this.size[0];
-					}
-					if (sz[1] < this.size[1]) {
-						sz[1] = this.size[1];
-					}
-					this.onResize?.(sz);
-					app.graph.setDirtyCanvas(true, false);
-				});
-			}
+        requestAnimationFrame(() => {
+          if (this) {
+            const sz = this.computeSize()
+            if (sz[0] < this.size[0]) {
+              sz[0] = this.size[0]
+            }
+            if (sz[1] < this.size[1]) {
+              sz[1] = this.size[1]
+            }
+            this.onResize?.(sz)
+            app.graph.setDirtyCanvas(true, false)
+          }
+        })
+      }

-			// When the node is executed we will be sent the input text, display this in the widget
-			const onExecuted = nodeType.prototype.onExecuted;
-			nodeType.prototype.onExecuted = function (message) {
-				onExecuted?.apply(this, arguments);
-				populate.call(this, message.text);
-			};
+      // When the node is executed we will be sent the input text, display this in the widget
+      const onExecuted = nodeType.prototype.onExecuted
+      nodeType.prototype.onExecuted = function (message) {
+        onExecuted?.apply(this, arguments)
+        // console.log('##onExecuted', this, message)
+        if (message.text) populate.call(this, message.text)
+      }

-			const onConfigure = nodeType.prototype.onConfigure;
-			nodeType.prototype.onConfigure = function () {
-				onConfigure?.apply(this, arguments);
-				if (this.widgets_values?.length) {
-          
-					populate.call(this, this.widgets_values);
-				}
-			};
+      const onConfigure = nodeType.prototype.onConfigure
+      nodeType.prototype.onConfigure = function () {
+        onConfigure?.apply(this, arguments)
+        if (this.widgets_values?.length) {
+          populate.call(this, this.widgets_values)
+        }
+      }

      this.serialize_widgets = true //需要保存参数
-
-		}
-
-
-	},
+    }
+  }
 })
@@ -1,12 +1,49 @@
 import { app } from '../../../scripts/app.js'
 import { api } from '../../../scripts/api.js'
-import { ComfyWidgets } from '../../../scripts/widgets.js'
+// import { ComfyWidgets } from '../../../scripts/widgets.js'
 import { $el } from '../../../scripts/ui.js'
+import { applyTextReplacements } from '../../../scripts/utils.js'

-async function uploadImage (blob,fileType='.svg') {
+import { loadExternalScript, get_position_style } from './common.js'
+
+function loadImageToCanvas (base64Image) {
+  var img = new Image()
+  var canvas = document.createElement('canvas')
+  var ctx = canvas.getContext('2d')
+  return new Promise((res, rej) => {
+    img.onload = function () {
+      // 等比例缩放图片
+      var width = img.width
+      var height = img.height
+      var max_width = 1024
+      if (width > max_width) {
+        height *= max_width / width
+        width = max_width
+      }
+
+      // 设置canvas尺寸
+      canvas.width = width
+      canvas.height = height
+
+      // 在canvas上绘制图片
+      ctx.drawImage(img, 0, 0, width, height)
+
+      // 将canvas转换为base64图片数据
+      var canvasData = canvas.toDataURL()
+      res(canvasData) //  canvas转换后的base64图片数据
+    }
+
+    img.src = base64Image
+  })
+}
+
+async function uploadImage (blob, fileType = '.svg', filename) {
  // const blob = await (await fetch(src)).blob();
  const body = new FormData()
-  body.append('image', new File([blob], new Date().getTime() + fileType))
+  body.append(
+    'image',
+    new File([blob], (filename || new Date().getTime()) + fileType)
+  )

  const resp = await api.fetchApi('/upload/image', {
    method: 'POST',
@@ -25,9 +62,11 @@ async function uploadImage (blob,fileType='.svg') {
  return src
 }

-function base64ToBlobFromURL(base64URL, contentType) {
-  return fetch(base64URL)
-    .then(response => response.blob());
+const base64Df =
+  'data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAwAAAAMCAYAAABWdVznAAAAAXNSR0IArs4c6QAAALZJREFUKFOFkLERwjAQBPdbgBkInECGaMLUQDsE0AkRVRAYWqAByxldPPOWHwnw4OBGye1p50UDSoA+W2ABLPN7i+C5dyC6R/uiAUXRQCs0bXoNIu4QPQzAxDKxHoALOrZcqtiyR/T6CXw7+3IGHhkYcy6BOR2izwT8LptG8rbMiCRAUb+CQ6WzQVb0SNOi5Z2/nX35DRyb/ENazhpWKoGwrpD6nICp5c2qogc4of+c7QcrhgF4Aa/aoAFHiL+RAAAAAElFTkSuQmCC'
+
+function base64ToBlobFromURL (base64URL, contentType) {
+  return fetch(base64URL).then(response => response.blob())
 }

 function getContentTypeFromBase64 (base64Data) {
@@ -51,37 +90,40 @@ function getContentTypeFromBase64 (base64Data) {
 // const blob = base64ToBlob(base64Data, contentType);
 // console.log(blob);

-function get_position_style (ctx, widget_width, y, node_height) {
-  const MARGIN = 4 // the margin around the html element
+// function get_position_style (ctx, widget_width, y, node_height) {
+//   const MARGIN = 4 // the margin around the html element

-  /* Create a transform that deals with all the scrolling and zooming */
-  const elRect = ctx.canvas.getBoundingClientRect()
-  const transform = new DOMMatrix()
-    .scaleSelf(
-      elRect.width / ctx.canvas.width,
-      elRect.height / ctx.canvas.height
-    )
-    .multiplySelf(ctx.getTransform())
-    .translateSelf(MARGIN, MARGIN + y)
+//   /* Create a transform that deals with all the scrolling and zooming */
+//   const elRect = ctx.canvas.getBoundingClientRect()
+//   const transform = new DOMMatrix()
+//     .scaleSelf(
+//       elRect.width / ctx.canvas.width,
+//       elRect.height / ctx.canvas.height
+//     )
+//     .multiplySelf(ctx.getTransform())
+//     .translateSelf(MARGIN, MARGIN + y)

-  return {
-    transformOrigin: '0 0',
-    transform: transform,
-    left: `0`,
-    top: `0`,
-    cursor: 'pointer',
-    position: 'absolute',
-    maxWidth: `${widget_width - MARGIN * 2}px`,
-    // maxHeight: `${node_height - MARGIN * 2}px`, // we're assuming we have the whole height of the node
-    width: `${widget_width - MARGIN * 2}px`,
-    // height: `${node_height * 0.3 - MARGIN * 2}px`,
-    // background: '#EEEEEE',
-    display: 'flex',
-    flexDirection: 'column',
-    // alignItems: 'center',
-    justifyContent: 'space-around'
-  }
-}
+//   return {
+//     transformOrigin: '0 0',
+//     transform: transform,
+//     left:
+//       document.querySelector('.comfy-menu').style.display === 'none'
+//         ? `60px`
+//         : `0`,
+//     top: `0`,
+//     cursor: 'pointer',
+//     position: 'absolute',
+//     maxWidth: `${widget_width - MARGIN * 2}px`,
+//     // maxHeight: `${node_height - MARGIN * 2}px`, // we're assuming we have the whole height of the node
+//     width: `${widget_width - MARGIN * 2}px`,
+//     // height: `${node_height * 0.3 - MARGIN * 2}px`,
+//     // background: '#EEEEEE',
+//     display: 'flex',
+//     flexDirection: 'column',
+//     // alignItems: 'center',
+//     justifyContent: 'space-around'
+//   }
+// }

 const getLocalData = key => {
  let data = {}
@@ -106,7 +148,7 @@ function createImage (url) {
  })
 }

-const parseImage = url => {
+const parseImageToBase64 = url => {
  return new Promise((res, rej) => {
    fetch(url)
      .then(response => response.blob())
@@ -126,6 +168,7 @@ const parseImage = url => {
 }

 const parseSvg = async svgContent => {
+  let scale = 2
  // 创建一个临时的DOM元素来解析SVG
  const tempContainer = document.createElement('div')
  tempContainer.innerHTML = svgContent
@@ -135,17 +178,18 @@ const parseSvg = async svgContent => {
  if (!svgElement) return
  // 获取SVG中 rect元素
  var rectElements = svgElement?.querySelectorAll('rect') || []
-
+  // console.log(rectElements,svgElement)
  // 定义一个数组来存储处理后的数据
  var data = []

  Array.from(rectElements, (rectElement, i) => {
    // 获取rect元素的属性值
-    var x = rectElement.getAttribute('x')
-    var y = rectElement.getAttribute('y')
-    var width = rectElement.getAttribute('width')
-    var height = rectElement.getAttribute('height')
-    if (x != undefined && y != undefined) {
+    var x = ~~(rectElement.getAttribute('x') || 0)
+    var y = ~~(rectElement.getAttribute('y') || 0)
+    var width = ~~rectElement.getAttribute('width')
+    var height = ~~rectElement.getAttribute('height')
+    // console.log('rectElements',rectElement,x,y,width,height)
+    if (x != undefined && y != undefined && width && height) {
      // 创建一个新的canvas元素
      var canvas = document.createElement('canvas')
      canvas.width = width
@@ -171,7 +215,8 @@ const parseSvg = async svgContent => {
        scale_option: 'width',
        image: base64,
        mask: base64,
-        type: 'base64'
+        type: 'base64',
+        _t: 'rect'
      }

      // 将处理后的数据添加到数组中
@@ -181,6 +226,20 @@ const parseSvg = async svgContent => {

  var svgWidth = svgElement.getAttribute('width')
  var svgHeight = svgElement.getAttribute('height')
+
+  if (!(svgWidth && svgHeight)) {
+    // viewBox
+    let viewBox = svgElement.viewBox.baseVal
+
+    svgWidth = viewBox.width
+    svgHeight = viewBox.height
+  } else {
+    try {
+      svgWidth = ~~svgWidth.replace('px', '')
+      svgHeight = ~~svgHeight.replace('px', '')
+    } catch (error) {}
+  }
+
  // 创建一个新的canvas元素
  var canvas = document.createElement('canvas')
  canvas.width = svgWidth
@@ -207,15 +266,35 @@ const parseSvg = async svgContent => {
    scale_option: 'width',
    image: base64,
    mask: base64,
-    type: 'base64'
+    type: 'base64',
+    _t: 'canvas'
  }
  data.push(rectData)

  // 打印处理后的数据
-  // console.log({ data, image: base64, svgElement })
+  console.log('layers', { data, image: base64, svgElement })
  return { data, image: base64, svgElement }
 }

+function exportModelViewerImage (
+  modelViewer,
+  width,
+  height,
+  format = 'image/png',
+  quality = 1.0
+) {
+  const canvas = document.createElement('canvas')
+  canvas.width = width
+  canvas.height = height
+  const context = canvas.getContext('2d')
+
+  return new Promise((resolve, reject) => {
+    context.drawImage(modelViewer, 0, 0, width, height)
+
+    resolve(canvas.toDataURL(format, quality))
+  })
+}
+
 app.registerExtension({
  name: 'Mixlab.image.SvgImage',
  async getCustomWidgets (app) {
@@ -270,7 +349,7 @@ app.registerExtension({
          draw (ctx, node, widget_width, y, widget_height) {
            Object.assign(
              this.div.style,
-              get_position_style(ctx, widget_width, 44, node.size[1])
+              get_position_style(ctx, widget_width, 44, node.size[1], 60)
            )
          }
        }
@@ -322,11 +401,17 @@ app.registerExtension({
              setLocalDataOfWin(key, dd)
              // console.log(this.id, ip.value.trim())

-              svgElement.style = `width: 90%;padding: 5%;`
+              svgElement.style = `width: 90%;padding: 5%;height: auto;`
              // 将提取的SVG元素显示在页面上

              svgContainer.innerHTML = ''
              svgContainer.appendChild(svgElement)
+              let h = ~~getComputedStyle(svgElement).height.replace('px', '')
+              if (that.size && that.size[1] < h) {
+                that.setSize([that.size[0], that.size[1] + h])
+                app.canvas.draw(true, true)
+              }
+              // console.log(that.size,~~getComputedStyle(svgElement).height.replace('px',''))

              uploadWidget.value = await uploadWidget.serializeValue()
            }
@@ -384,7 +469,7 @@ app.registerExtension({
      let svgStr = await dt.text()

      const { svgElement, data, image } = await parseSvg(svgStr)
-      svgElement.style = `width: 90%;padding: 5%;`
+      svgElement.style = `width: 90%;padding: 5%;height:auto`
      // 将提取的SVG元素显示在页面上

      widget.div.querySelector('.preview').innerHTML = ''
@@ -392,189 +477,152 @@ app.registerExtension({

      const uploadWidget = node.widgets.filter(w => w.name == 'upload')[0]
      uploadWidget.value = await uploadWidget.serializeValue()
-
-      // let h=~~getComputedStyle(widget.div).height.replace('px','');
-      // let w=~~getComputedStyle(widget.div).width.replace('px','');
-      // // console.log('svg', w,h,node.size)
-      // node.setSize([
-      //   w,h
-      // ])
-      // app.graph.setDirtyCanvas(true)
-
-      // console.log(node.widgets_values)
    }
  }
 })

-app.registerExtension({
-  name: 'Mixlab.image.3DImage',
-  async getCustomWidgets (app) {
-    return {
-      THREED (node, inputName, inputData, app) {
-        // console.log('##node', node, inputName, inputData)
-        const widget = {
-          type: inputData[0], // the type, CHEESE
-          name: inputName, // the name, slice
-          size: [128, 88], // a default size
-          draw (ctx, node, width, y) {},
-          computeSize (...args) {
-            return [128, 88] // a method to compute the current size of the widget
-          },
-          async serializeValue (nodeId, widgetIndex) {
-            let d = getLocalData('_mixlab_3d_image')
-            // console.log('serializeValue',d)
-            if (d) {
-              let url = d[node.id]
-              let base64 = await parseImage(url)
-
-              return JSON.parse(JSON.stringify({ image: base64 }))
-            } else {
-              return {}
-            }
-          }
-        }
-        node.addCustomWidget(widget)
-        return widget
-      }
+const createSelect = (imgDiv, select, opts, targetWidget, textWidget) => {
+  select.style.display = 'block'
+  let html = ''
+  let isMatch = false
+  for (const opt of opts) {
+    html += `<option value='${opt.keyword}' ${opt.selected ? 'selected' : ''}>${
+      opt.keyword
+    }</option>`
+    if (opt.selected) {
+      isMatch = true
+      imgDiv.src = opt.imgurl
+      // targetWidget.value = opt.keyword
    }
-  },
+  }
+  select.innerHTML = html
+  if (!isMatch) {
+    // targetWidget.value = opts[0].keyword
+    imgDiv.src = opts[0].imgurl
+  }

+  // 添加change事件监听器
+  select.addEventListener('change', async function () {
+    // 获取选中的选项的值
+    var selectedOption = select.options[select.selectedIndex].value
+    let t = opts.filter(opt => opt.keyword === selectedOption)[0]
+
+    targetWidget.value = await parseImageToBase64(t.imgurl)
+    imgDiv.src = targetWidget.value
+    textWidget.value = t.keyword
+  })
+  // console.log(select)
+}
+
+app.registerExtension({
+  name: 'Mixlab.prompt.ImagesPrompt_',
  async beforeRegisterNodeDef (nodeType, nodeData, app) {
-    if (nodeType.comfyClass == '3DImage') {
+    if (nodeType.comfyClass == 'ImagesPrompt_') {
      const orig_nodeCreated = nodeType.prototype.onNodeCreated
      nodeType.prototype.onNodeCreated = async function () {
        orig_nodeCreated?.apply(this, arguments)

-        const uploadWidget = this.widgets.filter(w => w.name == 'upload')[0]
-        console.log('3d nodeData', this.inputs)
+        const image_prompt = this.widgets.filter(
+          w => w.name == 'image_base64'
+        )[0]
+        const image_text = this.widgets.filter(w => w.name == 'text')[0]
+
+        const node = this

        const widget = {
          type: 'div',
-          name: 'upload-preview',
+          name: 'upload',
          draw (ctx, node, widget_width, y, widget_height) {
            Object.assign(
              this.div.style,
-              get_position_style(ctx, widget_width, 44, node.size[1])
+              get_position_style(ctx, widget_width, y, node.size[1], 36)
            )
          }
        }

        widget.div = $el('div', {})

-        document.body.appendChild(widget.div)
+        // console.log('image_prompt',image_prompt)
+        const img = new Image()
+        img.src = image_prompt?.value || base64Df
+        widget.div.appendChild(img)

-        const inputDiv = (key, placeholder, preview) => {
-          let div = document.createElement('div')
-          const ip = document.createElement('input')
-          ip.type = 'file'
-          ip.className = `${'comfy-multiline-input'} ${placeholder}`
-          div.style = `display: flex;
-          align-items: center; 
-          margin: 6px 8px;
-          margin-top: 0;`
-          ip.placeholder = placeholder
-          // ip.value = value
+        const btn = document.createElement('button')
+        btn.innerText = 'Upload Images JSON'

-          ip.style = `outline: none;
-          border: none;
-          padding: 4px;
-          width: 60%;cursor: pointer;
-          height: 32px;`
-          const label = document.createElement('label')
-          label.style = 'font-size: 10px;min-width:32px'
-          label.innerText = placeholder
-          div.appendChild(label)
-          div.appendChild(ip)
+        btn.style = `cursor: pointer;
+        font-weight: 300;
+        margin: 2px; 
+        color: var(--descrip-text);
+        background-color: var(--comfy-input-bg);
+        border-radius: 8px;
+        border-color: var(--border-color);
+        border-style: solid;height: 30px;min-width: 122px;
+       `

-          let that = this
+        const select = document.createElement('select')
+        select.style = `display:none;cursor: pointer;
+        font-weight: 300;
+        margin: 2px; 
+        color: var(--descrip-text);
+        background-color: var(--comfy-input-bg);
+        border-radius: 8px;
+        border-color: var(--border-color);
+        border-style: solid;height: 30px;min-width: 100px;
+       `
+        widget.select = select

-          ip.addEventListener('change', event => {
+        // const btn=document.createElement('button');
+        // btn.innerText='Upload'
+        btn.addEventListener('click', () => {
+          let inp = document.createElement('input')
+          inp.type = 'file'
+          inp.accept = '.json'
+          inp.click()
+          inp.addEventListener('change', event => {
+            // 获取选择的文件
+            // [{title,imageUrl}]
            const file = event.target.files[0]
+            this.title = file.name.split('.')[0]
+
+            // console.log(file.name.split('.')[0])
+            // 创建文件读取器
            const reader = new FileReader()

-            // 读取文件内容
-            reader.onload = async e => {
-              const fileURL = URL.createObjectURL(file)
-              // console.log('文件URL: ', fileURL)
-              let html = `<model-viewer 
-              alt="Neil Armstrong's Spacesuit from the Smithsonian Digitization Programs Office and National Air and Space Museum"
-               src="${fileURL}" 
-               ar 
-               shadow-intensity="1" 
-               camera-controls 
-               touch-action="pan-y">
-               
-               <div class="controls">
-                <div>Variant: <select class="variant"></select></div>
-                <div><button class="capture">Capture</button></div>
-              </div></model-viewer>`
+            // 定义读取完成事件的回调函数
+            reader.onload = async event => {
+              // 读取完成后的文本内容
+              const json = JSON.parse(event.target.result)
+              console.log(node, json)

-              preview.innerHTML = html
+              widget.value = JSON.stringify(json)

-              const modelViewerVariants = preview.querySelector('model-viewer')
-              const select = preview.querySelector('.variant')
-              const capture = preview.querySelector('.capture')
+              let img = widget.div.querySelector('img')

-              modelViewerVariants.addEventListener('load', () => {
-                const names = modelViewerVariants.availableVariants
-                for (const name of names) {
-                  const option = document.createElement('option')
-                  option.value = name
-                  option.textContent = name
-                  select.appendChild(option)
-                }
-                // Adds a default option.
-                const option = document.createElement('option')
-                option.value = 'default'
-                option.textContent = 'Default'
-                select.appendChild(option)
-              })
+              createSelect(img, select, json, image_prompt, image_text)

-              select.addEventListener('input', event => {
-                modelViewerVariants.variantName =
-                  event.target.value === 'default' ? null : event.target.value
-              })
+              image_prompt.value = await parseImageToBase64(json[0].imgurl)
+              image_text.value = json[0].keyword

-              capture.addEventListener('click', async () => {
-                let base64Data = modelViewerVariants.toDataURL()
+              if (img) {
+                img.src = image_prompt.value
+              }

-                const contentType = getContentTypeFromBase64(base64Data)
-
-                const blob =await base64ToBlobFromURL(base64Data, contentType)
-
-                //  const fileBlob = new Blob([e.target.result], { type: file.type });
-                let url = await uploadImage(blob,'.png')
-                console.log(url)
-
-                let dd = getLocalData(key)
-                dd[that.id] = url
-
-                setLocalDataOfWin(key, dd)
-              })
-
-              uploadWidget.value = await uploadWidget.serializeValue()
+              inp.remove()
            }

-            // 以文本形式读取文件
-            reader.readAsDataURL(file)
+            // 以文本方式读取文件
+            reader.readAsText(file)
          })
-          return div
-        }
+        })

-        let preview = document.createElement('div')
-        preview.className = 'preview'
-        preview.style = `background:#eee;margin-top: 12px;`
-
-        let upload = inputDiv('_mixlab_3d_image', '3D Model', preview)
-
-        widget.div.appendChild(upload)
-        widget.div.appendChild(preview)
+        widget.div.appendChild(btn)
+        widget.div.appendChild(select)
+        document.body.appendChild(widget.div)
        this.addCustomWidget(widget)

        const onRemoved = this.onRemoved
        this.onRemoved = () => {
-          upload.remove()
-          preview.remove()
          widget.div.remove()
          return onRemoved?.()
        }
@@ -588,40 +636,392 @@ app.registerExtension({
    }
  },
  async loadedGraphNode (node, app) {
-    // Fires every time a node is constructed
-    // You can modify widgets/add handlers/etc here
-    const sleep = (t = 1000) => {
-      return new Promise((res, rej) => {
-        setTimeout(() => res(1), t)
-      })
-    }
-    if (node.type === '3DImage') {
-      // await sleep(0)
-      let widget = node.widgets.filter(w => w.name === 'upload-preview')[0]
+    if (node.type === 'ImagesPrompt_') {
+      try {
+        let prompt = node.widgets.filter(w => w.name === 'image_base64')[0]
+        let text = node.widgets.filter(w => w.name === 'text')[0]
+        let uploadWidget = node.widgets.filter(w => w.name == 'upload')[0]
+        // console.log('##prompt',prompt.value)
+        let img = uploadWidget.div.querySelector('img')
+        let json = JSON.parse(uploadWidget.value)

-      let dd = getLocalData('_mixlab_3d_image')
+        for (let index = 0; index < json.length; index++) {
+          const j = json[index]
+          let base64 = await parseImageToBase64(j.imgurl)
+          if (base64 === prompt.value) {
+            json[index].selected = true
+          }
+        }

-      let id = node.id
-      console.log('3dImage load', node.widgets[0], node.widgets)
-      if (!dd[id]) return
+        if (json && json[0]) {
+          uploadWidget.select.style.display = 'block'
+          createSelect(img, uploadWidget.select, json, prompt, text)
+        }
+      } catch (error) {}
+    }
+  }
+})

-      let url = dd[id]
-      // let base64 = await parseImage(url)
+const createInputImageForBatch = (base64, widget) => {
+  let im = new Image()
+  im.src = base64
+  im.style = `width: 88px;`

-      widget.div.querySelector('.preview').innerHTML = `<img src="${url}"/>` 
+  im.addEventListener('click', e => {
+    let newValue = []
+    let items = widget.value?.base64 || []
+    for (const v of items) {
+      if (v != base64) newValue.push(v)
+    }
+    widget.value.base64 = newValue
+    im.remove()
+  })

-      const uploadWidget = node.widgets.filter(w => w.name == 'upload')[0]
-      uploadWidget.value = await uploadWidget.serializeValue()
+  return im
+}

-      // let h=~~getComputedStyle(widget.div).height.replace('px','');
-      // let w=~~getComputedStyle(widget.div).width.replace('px','');
-      // // console.log('svg', w,h,node.size)
-      // node.setSize([
-      //   w,h
-      // ])
-      // app.graph.setDirtyCanvas(true)
+// 添加新图片
+const addBase64ToWidgetForLoadImagesToBatch = (
+  base64,
+  imagesWidget,
+  imagesDiv
+) => {
+  if (!imagesWidget.value.base64) imagesWidget.value.base64 = []
+  imagesWidget.value.base64.push(base64)
+  let im = createInputImageForBatch(base64, imagesWidget)
+  imagesDiv.appendChild(im)
+}

-      // console.log(node.widgets_values)
+app.registerExtension({
+  name: 'Mixlab.Comfy.LoadImagesToBatch',
+  async getCustomWidgets (app) {
+    return {
+      IMAGEBASE64 (node, inputName, inputData, app) {
+        // console.log('##node', node)
+        const widget = {
+          value: {
+            base64: []
+          }, // 不能[x,x,x]
+          type: inputData[0], // the type
+          name: inputName, // the name, slice
+          size: [128, 32], // a default size
+          draw (ctx, node, width, y) {},
+          computeSize (...args) {
+            return [128, 32] // a method to compute the current size of the widget
+          }
+          // serializeValue (nodeId, widgetIndex) {
+          //   return widget.value
+          // },
+        }
+        //  widget.something = something;          // maybe adds stuff to it
+        node.addCustomWidget(widget) // adds it to the node
+        return widget // and returns it.
+      }
+    }
+  },
+
+  async beforeRegisterNodeDef (nodeType, nodeData, app) {
+    if (nodeType.comfyClass == 'LoadImagesToBatch') {
+      const orig_nodeCreated = nodeType.prototype.onNodeCreated
+
+      nodeType.prototype.onNodeCreated = function () {
+        orig_nodeCreated?.apply(this, arguments)
+
+        let imagesWidget = this.widgets.filter(w => w.name == 'images')[0]
+
+        const widget = {
+          type: 'div',
+          name: 'image_base64',
+          draw (ctx, node, widget_width, y, widget_height) {
+            Object.assign(this.div.style, {
+              ...get_position_style(ctx, widget_width, y, node.size[1], 72),
+              top: `${widget_height}px`
+            })
+          },
+          serialize: false
+        }
+
+        widget.div = $el('div', {})
+
+        document.body.appendChild(widget.div)
+
+        let imagePreview = document.createElement('div')
+        let imagesDiv = document.createElement('div') //显示图片
+        imagesDiv.className = 'images_preview'
+        imagesDiv.style = `width: calc(100% - 14px);
+        display: flex;
+        flex-wrap: wrap;
+        padding: 7px;    justify-content: space-between;
+        align-items: center;`
+
+        let inputImage = document.createElement('input')
+        inputImage.type = 'file'
+        inputImage.style.display = 'none'
+        inputImage.addEventListener('change', e => {
+          e.preventDefault()
+          const file = e.target.files[0]
+          const reader = new FileReader()
+          reader.onload = async event => {
+            let base64 = event.target.result
+            //压缩图片，控制1024以内
+            base64 = await loadImageToCanvas(base64)
+            // console.log(base64)
+            if (!imagesWidget.value) imagesWidget.value = { base64: [] }
+            addBase64ToWidgetForLoadImagesToBatch(
+              base64,
+              imagesWidget,
+              imagesDiv
+            )
+          }
+          reader.readAsDataURL(file)
+        })
+
+        // 如果是复制的，有数据 , 这个不生效，取不到数据， 需要在nodeCreated里获取
+        // console.log('#LoadImagesToBatch', imagesWidget.value?.base64)
+
+        const btn = document.createElement('button')
+        btn.innerText = 'Upload Image'
+
+        btn.style = `cursor: pointer;
+        font-weight: 300;
+        margin: 2px; 
+        color: var(--descrip-text);
+        background-color: var(--comfy-input-bg);
+        border-radius: 8px;
+        border-color: var(--border-color);
+        border-style: solid;height: 30px;min-width: 122px;
+       `
+
+        btn.addEventListener('click', e => {
+          e.preventDefault()
+          inputImage.click()
+        })
+
+        widget.div.appendChild(imagePreview)
+        imagePreview.appendChild(imagesDiv)
+        imagePreview.appendChild(btn)
+        imagePreview.appendChild(inputImage)
+
+        this.addCustomWidget(widget)
+
+        // document.addEventListener('wheel', handleMouseWheel)
+
+        const onRemoved = this.onRemoved
+        this.onRemoved = () => {
+          inputImage.remove()
+          widget.div.remove()
+          try {
+            // document.removeEventListener('wheel', handleMouseWheel)
+          } catch (error) {
+            console.log(error)
+          }
+
+          return onRemoved?.()
+        }
+
+        this.serialize_widgets = true //需要保存参数
+      }
+    }
+
+    if (nodeData.name === 'SaveImageAndMetadata_') {
+      const onNodeCreated = nodeType.prototype.onNodeCreated
+      // /web/extensions/core/saveImageExtraOutput.js
+      nodeType.prototype.onNodeCreated = function () {
+        const r = onNodeCreated
+          ? onNodeCreated.apply(this, arguments)
+          : undefined
+        const widget = this.widgets.find(w => w.name === 'filename_prefix')
+        widget.serializeValue = () => {
+          return applyTextReplacements(app, widget.value)
+        }
+
+        return r
+      }
+
+      const onExecuted = nodeType.prototype.onExecuted
+      nodeType.prototype.onExecuted = function (message) {
+        onExecuted?.apply(this, arguments)
+        console.log('##onExecuted', this, message)
+        //TODO 是否 保存base64
+        if (message.base64) {
+          if (Array.isArray(message.base64)) {
+          }
+        }
+      }
+    }
+  },
+
+  async loadedGraphNode (node, app) {
+    if (node.type === 'LoadImagesToBatch') {
+      let imagesWidget = node.widgets.filter(w => w.name === 'images')[0]
+      let imagePreview = node.widgets.filter(w => w.name == 'image_base64')[0]
+      // console.log('#LoadImagesToBatch', imagesWidget.value?.base64)
+      let imagesDiv = imagePreview.div.querySelector('.images_preview')
+      imagesDiv.innerHTML = ''
+      for (const d of imagesWidget.value?.base64 || []) {
+        let im = createInputImageForBatch(d, imagesWidget)
+        imagesDiv.appendChild(im)
+      }
+    }
+  },
+  nodeCreated (node, app) {
+    //数据延迟？？
+    setTimeout(() => {
+      // console.log('#LoadImagesToBatch', node.type)
+      if (node.type === 'LoadImagesToBatch') {
+        let imagesWidget = node.widgets.filter(w => w.name === 'images')[0]
+        let imagePreview = node.widgets.filter(w => w.name == 'image_base64')[0]
+
+        let imagesDiv = imagePreview?.div?.querySelector('.images_preview')
+        imagesDiv.innerHTML = ''
+        for (const d of imagesWidget.value?.base64 || []) {
+          let im = createInputImageForBatch(d, imagesWidget)
+          imagesDiv.appendChild(im)
+        }
+      }
+    }, 1000)
+  }
+})
+
+// 如何引入css
+app.registerExtension({
+  name: 'Mixlab.output.ComparingTwoFrames_',
+  init () {
+    loadExternalScript('/mixlab/app/lib/juxtapose.min.js')
+
+    $el('link', {
+      rel: 'stylesheet',
+      href: '/mixlab/app/lib/juxtapose.css',
+      parent: document.head
+    })
+
+    $el('style', {
+      textContent: `
+      .juxtapose-name{
+        display: none!important; 
+      }
+			`,
+      parent: document.body
+    })
+  },
+  async beforeRegisterNodeDef (nodeType, nodeData, app) {
+    if (nodeType.comfyClass == 'ComparingTwoFrames_') {
+      const onNodeCreated = nodeType.prototype.onNodeCreated
+      nodeType.prototype.onNodeCreated = function () {
+        const r = onNodeCreated
+          ? onNodeCreated.apply(this, arguments)
+          : undefined
+
+        this.size = [400, this.size[1]]
+        console.log('##onNodeCreated', this)
+        const widget = {
+          type: 'div',
+          name: 'preview',
+          draw (ctx, node, widget_width, y, widget_height) {
+            let s = get_position_style(ctx, widget_width, 44, node.size[1], 36)
+            delete s.height
+
+            Object.assign(this.div.style, s)
+          },
+          serialize: false
+        }
+
+        widget.div = $el('div', {})
+
+        document.body.appendChild(widget.div)
+        this.addCustomWidget(widget)
+        this.serialize_widgets = true //需要保存参数
+
+        const onRemoved = this.onRemoved
+        this.onRemoved = () => {
+          widget.div.remove()
+          return onRemoved?.()
+        }
+
+        return r
+      }
+
+      const onExecuted = nodeType.prototype.onExecuted
+      nodeType.prototype.onExecuted = function (message) {
+        onExecuted?.apply(this, arguments)
+        console.log('##onExecuted', this, message)
+
+        this.widgets[0].div.id = 'mix_comparingtowframes_' + this.id
+
+        let after_image = message.after_images[0]
+        let before_image = message.before_images[0]
+
+        after_image = `${window.location.protocol}//${
+          window.location.hostname
+        }:${window.location.port}/view?filename=${encodeURIComponent(
+          after_image.filename
+        )}&type=${after_image.type}&subfolder=${encodeURIComponent(
+          after_image.subfolder
+        )}&t=${+new Date()}`
+
+        before_image = `${window.location.protocol}//${
+          window.location.hostname
+        }:${window.location.port}/view?filename=${encodeURIComponent(
+          before_image.filename
+        )}&type=${before_image.type}&subfolder=${encodeURIComponent(
+          before_image.subfolder
+        )}&t=${+new Date()}`
+
+        this.widgets[0].div.innerHTML = ''
+
+        let slider = new juxtapose.JXSlider(
+          '#mix_comparingtowframes_' + this.id,
+          [
+            {
+              src: before_image,
+              label: 'Before'
+            },
+            {
+              src: after_image,
+              label: 'After'
+            }
+          ],
+          {
+            animate: true,
+            showLabels: true,
+            showCredits: false,
+            startingPosition: '50%',
+            makeResponsive: false
+          }
+        )
+
+        this.widgets_values = [
+          {
+            src: before_image,
+            label: 'Before'
+          },
+          {
+            src: after_image,
+            label: 'After'
+          }
+        ]
+        this.size = [this.size[0], 300]
+      }
+    }
+  },
+  async loadedGraphNode (node, app) {
+    // console.log('##loadedGraphNode', node)
+    if (node.type === 'ComparingTwoFrames_') {
+      // node.widgets[0].div.id = 'mix_comparingtowframes_' + node.id
+      // if (node.widgets_values && node.widgets_values[0]) {
+      //   node.widgets[0].div.innerHTML = ''
+      //   let slider = new juxtapose.JXSlider(
+      //     '#mix_comparingtowframes_' + node.id,
+      //     node.widgets_values,
+      //     {
+      //       animate: true,
+      //       showLabels: true,
+      //       showCredits: false,
+      //       startingPosition: '50%',
+      //       makeResponsive: false
+      //     }
+      //   )
+      // }
    }
  }
 })
@@ -1,8 +1,70 @@
 import { app } from '../../../scripts/app.js'
-// import { api } from '../../../scripts/api.js'
+import { api } from '../../../scripts/api.js'
 import { ComfyWidgets } from '../../../scripts/widgets.js'
 import { $el } from '../../../scripts/ui.js'

+function downloadJsonFile (jsonData, fileName = 'grid.json') {
+  const dataString = JSON.stringify(jsonData)
+  const blob = new Blob([dataString], { type: 'application/json' })
+  const url = URL.createObjectURL(blob)
+
+  const link = document.createElement('a')
+  link.href = url
+  link.download = fileName
+  link.click()
+
+  // 释放URL对象
+  setTimeout(() => {
+    URL.revokeObjectURL(url)
+  }, 0)
+}
+
+function createSelectWithOptions (options) {
+  const select = document.createElement('select')
+
+  options.forEach(option => {
+    const optionElement = document.createElement('option')
+    optionElement.text = option
+    optionElement.value = option
+    select.appendChild(optionElement)
+  })
+
+  select.style = `cursor: pointer;
+  font-weight: 300; 
+  height: 30px;
+  min-width: 122px;
+  position: absolute;
+  top: 24px;
+  left: 88px;
+  z-index: 999999999999999;
+       `
+
+  return select
+}
+
+function drawCanvasWithText (w, h, tag, color = 'rgba(255,255,255,0.4)') {
+  const canvas = document.createElement('canvas')
+  const ctx = canvas.getContext('2d')
+
+  // 设置画布大小
+  canvas.width = w
+  canvas.height = h
+
+  // 绘制白色背景
+  ctx.fillStyle = color
+  ctx.fillRect(0, 0, canvas.width, canvas.height)
+
+  // 绘制文字
+  ctx.fillStyle = '#000000'
+  ctx.font = '20px Arial'
+  ctx.fillText(tag, 50, 50)
+
+  // 导出为Base64
+  const base64 = canvas.toDataURL()
+
+  return base64
+}
+
 function get_position_style (ctx, widget_width, y, node_height) {
  const MARGIN = 4 // the margin around the html element

@@ -19,7 +81,10 @@ function get_position_style (ctx, widget_width, y, node_height) {
  return {
    transformOrigin: '0 0',
    transform: transform,
-    left: `0`,
+    left:
+    document.querySelector('.comfy-menu').style.display === 'none'
+      ? `60px`
+      : `0`,
    top: `0`,
    cursor: 'pointer',
    position: 'absolute',
@@ -63,17 +128,18 @@ const parseSvg = async svgContent => {
  if (!svgElement) return
  // 获取SVG中 rect元素
  var rectElements = svgElement?.querySelectorAll('rect') || []
-
+  // console.log(rectElements,svgElement)
  // 定义一个数组来存储处理后的数据
  var data = []

  Array.from(rectElements, (rectElement, i) => {
    // 获取rect元素的属性值
-    var x = rectElement.getAttribute('x')
-    var y = rectElement.getAttribute('y')
-    var width = rectElement.getAttribute('width')
-    var height = rectElement.getAttribute('height')
-    if (x != undefined && y != undefined) {
+    var x = ~~(rectElement.getAttribute('x') || 0)
+    var y = ~~(rectElement.getAttribute('y') || 0)
+    var width = ~~rectElement.getAttribute('width')
+    var height = ~~rectElement.getAttribute('height')
+    // console.log('rectElements',rectElement,x,y,width,height)
+    if (x != undefined && y != undefined && width && height) {
      // 创建一个新的canvas元素
      var canvas = document.createElement('canvas')
      canvas.width = width
@@ -99,7 +165,8 @@ const parseSvg = async svgContent => {
        scale_option: 'width',
        image: base64,
        mask: base64,
-        type: 'base64'
+        type: 'base64',
+        _t: 'rect'
      }

      // 将处理后的数据添加到数组中
@@ -109,6 +176,15 @@ const parseSvg = async svgContent => {

  var svgWidth = svgElement.getAttribute('width')
  var svgHeight = svgElement.getAttribute('height')
+
+  if (!(svgWidth && svgHeight)) {
+    // viewBox
+    let viewBox = svgElement.viewBox.baseVal
+
+    svgWidth = viewBox.width
+    svgHeight = viewBox.height
+  }
+
  // 创建一个新的canvas元素
  var canvas = document.createElement('canvas')
  canvas.width = svgWidth
@@ -135,15 +211,400 @@ const parseSvg = async svgContent => {
    scale_option: 'width',
    image: base64,
    mask: base64,
-    type: 'base64'
+    type: 'base64',
+    _t: 'canvas'
  }
  data.push(rectData)

  // 打印处理后的数据
-  console.log({ data, image: base64, svgElement })
+  console.log('layers', { data, image: base64, svgElement })
  return { data, image: base64, svgElement }
 }

+function findImages (nodeId) {
+  // 检查当前节点是否有 imgs 字段
+  const n = app.graph.getNodeById(nodeId)
+  if (n.imgs) {
+    return n.imgs
+  }
+
+  // 检查当前节点的 inputs 是否有 image 字段
+  if (n.inputs) {
+    for (let i = 0; i < n.inputs.length; i++) {
+      if (n.inputs[i].name === 'image' || n.inputs[i].name === 'images') {
+        // 获取新的 nodeId，并递归调用 findImages 函数
+        var linkId = n.inputs[i]?.link
+        var origin_id = app.graph.links[linkId].origin_id
+        return findImages(origin_id)
+      }
+    }
+  }
+
+  // 如果没有找到 imgs 字段或者 image 字段，则返回 null
+  return null
+}
+
+async function setArea (cw, ch, topBase64, base64, data, fn) {
+  let displayHeight = Math.round(window.screen.availHeight * 0.8)
+  let div = document.createElement('div')
+  div.innerHTML = `
+    <div id='ml_overlay' style='position: absolute;top:0;background: #251f1fc4;
+    height: 100vh;
+    z-index:999999;
+    width: 100%;'>
+      <img id='ml_video' style='position: absolute; 
+      height: ${displayHeight}px;user-select: none; 
+      -webkit-user-drag: none;
+      outline: 2px solid #eaeaea;
+      box-shadow: 8px 9px 17px #575757;' />
+      <div id='ml_selection' style='position: absolute; 
+      border: 2px dashed red;
+      pointer-events: none;
+      background-image: url("${topBase64}");
+            background-repeat: no-repeat;
+            background-size: cover;
+            '></div>
+      <div class="mx_close"> X </div>
+    </div>`
+  // document.body.querySelector('#ml_overlay')
+  document.body.appendChild(div)
+
+  // let canvas = document.createElement('canvas')
+  // canvas.width = cw
+  // canvas.height = ch
+
+  let img = div.querySelector('#ml_video')
+  // let overlay = div.querySelector('#ml_overlay')
+  let selection = div.querySelector('#ml_selection')
+  let close = div.querySelector('.mx_close')
+  let startX, startY, endX, endY
+  let start = false
+  let setDone = false
+  // Set video source
+  img.src = base64
+  // canvas.toDataURL();
+  close.style = `cursor: pointer;
+  position: fixed;
+  left: 12px;
+  top: 12px;
+  z-index: 99999999;
+  background: black;
+  width: 44px;
+  height: 44px;
+  text-align: center;
+  line-height: 44px;`
+
+  // init area
+  // const data = getSetAreaData()
+  let x = 0,
+    y = 0,
+    width = (cw * displayHeight) / ch,
+    height = displayHeight
+
+  let imgWidth = cw
+  let imgHeight = ch
+
+  if (data && data.width > 0 && data.height > 0) {
+    // 相同尺寸窗口，恢复选区
+    x = (width * data.x) / imgWidth
+    y = (height * data.y) / imgHeight
+    width = (width * data.width) / imgWidth
+    height = (height * data.height) / imgHeight
+  }
+
+  selection.style.left = x + 'px'
+  selection.style.top = y + 'px'
+  selection.style.width = width + 'px'
+  selection.style.height = height + 'px'
+
+  // Add mouse events
+  img.addEventListener('mousedown', startSelection)
+  img.addEventListener('mousemove', updateSelection)
+  img.addEventListener('mouseup', endSelection)
+
+  const removeDiv = () => {
+    div.remove()
+    close.removeEventListener('click', removeDiv)
+    img.removeEventListener('mousedown', startSelection)
+    img.removeEventListener('mousemove', updateSelection)
+    img.removeEventListener('mouseup', endSelection)
+    img.removeEventListener('mousedown', setDoneCheck)
+  }
+  close.addEventListener('click', removeDiv)
+
+  const setDoneCheck = event => {
+    console.log(setDone)
+    if (setDone) {
+      img.addEventListener('mousedown', startSelection)
+      img.addEventListener('mousemove', updateSelection)
+      img.addEventListener('mouseup', endSelection)
+      setDone = false
+      start = false
+      startX = event.clientX
+      startY = event.clientY
+    }
+  }
+  img.addEventListener('mousedown', setDoneCheck)
+
+  function remove () {
+    img.removeEventListener('mousedown', startSelection)
+    img.removeEventListener('mousemove', updateSelection)
+    img.removeEventListener('mouseup', endSelection)
+    setDone = true
+    // div.remove()
+  }
+
+  function startSelection (event) {
+    if (start == false) {
+      startX = event.clientX
+      startY = event.clientY
+      updateSelection(event)
+      start = true
+    } else {
+    }
+  }
+
+  function updateSelection (event) {
+    endX = event.clientX
+    endY = event.clientY
+
+    // Calculate width, height, and coordinates
+    let width = Math.abs(endX - startX)
+    let height = Math.abs(endY - startY)
+    let left = Math.min(startX, endX)
+    let top = Math.min(startY, endY)
+
+    // Set selection style
+    selection.style.left = left + 'px'
+    selection.style.top = top + 'px'
+    selection.style.width = width + 'px'
+    selection.style.height = height + 'px'
+  }
+
+  function endSelection (event) {
+    endX = event.clientX
+    endY = event.clientY
+
+    // 获取img元素的真实宽度和高度
+    let imgWidth = img.naturalWidth
+    let imgHeight = img.naturalHeight
+
+    // 换算起始坐标
+    let realStartX = (startX / img.offsetWidth) * imgWidth
+    let realStartY = (startY / img.offsetHeight) * imgHeight
+
+    // 换算起始坐标
+    let realEndX = (endX / img.offsetWidth) * imgWidth
+    let realEndY = (endY / img.offsetHeight) * imgHeight
+
+    startX = realStartX
+    startY = realStartY
+    endX = realEndX
+    endY = realEndY
+    // Calculate width, height, and coordinates
+    let width = Math.round(Math.abs(endX - startX))
+    let height = Math.round(Math.abs(endY - startY))
+    let left = Math.round(Math.min(startX, endX))
+    let top = Math.round(Math.min(startY, endY))
+
+    if (width <= 0 && height <= 0) return remove()
+
+    if (fn) fn(left, top, width, height)
+
+    remove()
+  }
+}
+
+async function setAreaTags (cw, ch, grids, fn) {
+  let base64 = drawCanvasWithText(cw, ch, '', 'white')
+  let displayHeight = Math.round(window.screen.availHeight * 0.8)
+  let div = document.createElement('div')
+  div.innerHTML = `
+    <div id='ml_overlay' style='position: absolute;top:0;background: #251f1fc4;
+    height: 100vh;
+    z-index:999999;
+    width: 100%;'>
+      <img id='ml_video' style='position: absolute; 
+      height: ${displayHeight}px;user-select: none; 
+      -webkit-user-drag: none;
+      outline: 2px solid #eaeaea;
+      box-shadow: 8px 9px 17px #575757;' />
+      ${Array.from(grids, g => {
+        const { label: tag, grid } = g
+        const [dx, dy, dw, dh] = grid
+        const base64Data = drawCanvasWithText(dw, dh, tag)
+
+        let x = 0,
+          y = 0,
+          width = (cw * displayHeight) / ch,
+          height = displayHeight
+
+        let imgWidth = cw
+        let imgHeight = ch
+
+        if (dw > 0 && dh > 0) {
+          // 相同尺寸窗口，恢复选区
+          x = (width * dx) / imgWidth
+          y = (height * dy) / imgHeight
+          width = (width * dw) / imgWidth
+          height = (height * dh) / imgHeight
+        }
+
+        return `<div class='ml_selection' 
+          data-tag="${tag}"
+          style='position:absolute; 
+          border: 2px dashed red;
+          pointer-events: none;
+          background-image: url("${base64Data}");
+                background-repeat: no-repeat;
+                background-size: cover;
+                left:${x}px;
+                top:${y}px;
+                width:${width}px;
+                height:${height}px;
+                '></div>`
+      })}
+      <div class="mx_close"> X </div>
+    </div>`
+  // document.body.querySelector('#ml_overlay')
+  document.body.appendChild(div)
+
+  const tags = Array.from(grids, g => g.label)
+  let select = createSelectWithOptions(tags)
+  document.body.appendChild(select)
+
+  let img = div.querySelector('#ml_video')
+  // let overlay = div.querySelector('#ml_overlay')
+  let selections = [...div.querySelectorAll('.ml_selection')]
+
+  let selection = selections.filter(
+    s => s.getAttribute('data-tag') === select.value
+  )[0]
+
+  select.addEventListener('change', e => {
+    selection = selections.filter(
+      s => s.getAttribute('data-tag') === select.value
+    )[0]
+  })
+
+  // console.log(select.value,selection)
+  let close = div.querySelector('.mx_close')
+  let startX, startY, endX, endY
+  let start = false
+  let setDone = false
+  // Set video source
+  img.src = base64
+  // canvas.toDataURL();
+  close.style = `cursor: pointer;
+  position: fixed;
+  left: 12px;
+  top: 12px;
+  z-index: 99999999;
+  background: black;
+  width: 44px;
+  height: 44px;
+  text-align: center;
+  line-height: 44px;`
+
+  // Add mouse events
+  img.addEventListener('mousedown', startSelection)
+  img.addEventListener('mousemove', updateSelection)
+  img.addEventListener('mouseup', endSelection)
+
+  const removeDiv = () => {
+    div.remove()
+    select?.remove()
+    close.removeEventListener('click', removeDiv)
+    img.removeEventListener('mousedown', startSelection)
+    img.removeEventListener('mousemove', updateSelection)
+    img.removeEventListener('mouseup', endSelection)
+    img.removeEventListener('mousedown', setDoneCheck)
+  }
+  close.addEventListener('click', removeDiv)
+
+  const setDoneCheck = event => {
+    console.log(setDone)
+    if (setDone) {
+      img.addEventListener('mousedown', startSelection)
+      img.addEventListener('mousemove', updateSelection)
+      img.addEventListener('mouseup', endSelection)
+      setDone = false
+      start = false
+      startX = event.clientX
+      startY = event.clientY
+    }
+  }
+  img.addEventListener('mousedown', setDoneCheck)
+
+  function remove () {
+    img.removeEventListener('mousedown', startSelection)
+    img.removeEventListener('mousemove', updateSelection)
+    img.removeEventListener('mouseup', endSelection)
+    setDone = true
+    // select?.remove()
+  }
+
+  function startSelection (event) {
+    if (start == false) {
+      startX = event.clientX
+      startY = event.clientY
+      updateSelection(event)
+      start = true
+    } else {
+    }
+  }
+
+  function updateSelection (event) {
+    endX = event.clientX
+    endY = event.clientY
+
+    // Calculate width, height, and coordinates
+    let width = Math.abs(endX - startX)
+    let height = Math.abs(endY - startY)
+    let left = Math.min(startX, endX)
+    let top = Math.min(startY, endY)
+
+    // Set selection style
+    selection.style.left = left + 'px'
+    selection.style.top = top + 'px'
+    selection.style.width = width + 'px'
+    selection.style.height = height + 'px'
+  }
+
+  function endSelection (event) {
+    endX = event.clientX
+    endY = event.clientY
+
+    // 获取img元素的真实宽度和高度
+    let imgWidth = img.naturalWidth
+    let imgHeight = img.naturalHeight
+
+    // 换算起始坐标
+    let realStartX = (startX / img.offsetWidth) * imgWidth
+    let realStartY = (startY / img.offsetHeight) * imgHeight
+
+    // 换算起始坐标
+    let realEndX = (endX / img.offsetWidth) * imgWidth
+    let realEndY = (endY / img.offsetHeight) * imgHeight
+
+    startX = realStartX
+    startY = realStartY
+    endX = realEndX
+    endY = realEndY
+    // Calculate width, height, and coordinates
+    let width = Math.round(Math.abs(endX - startX))
+    let height = Math.round(Math.abs(endY - startY))
+    let left = Math.round(Math.min(startX, endX))
+    let top = Math.round(Math.min(startY, endY))
+
+    if (width <= 0 && height <= 0) return remove()
+
+    if (!!fn) fn(select.value, left, top, width, height)
+
+    remove()
+  }
+}
+
 app.registerExtension({
  name: 'Mixlab.layer.ShowLayer',
  async getCustomWidgets (app) {
@@ -187,8 +648,7 @@ app.registerExtension({
        const findNode = nodeId => {
          let node = app.graph._nodes_by_id[nodeId]
          if (node?.type == 'Reroute') {
-            
-            let linkId =node.inputs.filter(i=>i.type=='*')[0].link
+            let linkId = node.inputs.filter(i => i.type == '*')[0].link
            nodeId = app.graph.links.filter(link => link.id == linkId)[0]
              ?.origin_id
            return findNode(nodeId)
@@ -199,11 +659,17 @@ app.registerExtension({

        // 获取layers数据
        const getLayers = async () => {
+          console.log(
+            'getLayers1',
+            this.inputs.filter(ip => ip.name === 'layers')
+          )
          let linkId = this.inputs.filter(ip => ip.name === 'layers')[0].link
-          let nodeId = app.graph.links.filter(link => link.id == linkId)[0]
+          let nodeId = app.graph.links?.filter(link => link.id == linkId)[0]
            ?.origin_id

-          nodeId = findNode(nodeId)
+          if (nodeId) {
+            nodeId = findNode(nodeId)
+          }

          //   let node = app.graph._nodes_by_id[nodeId]
          //   if (node?.type == 'Reroute') {
@@ -211,15 +677,18 @@ app.registerExtension({
          //     nodeId = app.graph.links.filter(link => link.id == linkId)[0]
          //       ?.origin_id
          //   }
-         
+
          let d = getLocalData('_mixlab_svg_image')
-          console.log('test',d[nodeId])
+          console.log('test', d[nodeId])

          if (d[nodeId]) {
            let url = d[nodeId]
            let dt = await fetch(url)
+
            let svgStr = await dt.text()
+
            const { data } = (await parseSvg(svgStr)) || {}
+            console.log('fetch', data)
            return data
          } else {
            return []
@@ -331,3 +800,399 @@ app.registerExtension({
    }
  }
 })
+
+app.registerExtension({
+  name: 'Mixlab.layer.NewLayer',
+  async beforeRegisterNodeDef (nodeType, nodeData, app) {
+    if (nodeData.name === 'NewLayer') {
+      const orig_nodeCreated = nodeType.prototype.onNodeCreated
+      nodeType.prototype.onNodeCreated = async function () {
+        orig_nodeCreated?.apply(this, arguments)
+
+        let b = this.widgets.filter(w => w.type === 'button')[0]
+        // const [w, h, base64] = canvas
+
+        if (!b) {
+          const updateValue = (x1, y1, w1, h1) => {
+            if (this.widgets) {
+              for (const widget of this.widgets) {
+                if (widget.name === 'x') {
+                  widget.value = x1
+                }
+                if (widget.name === 'y') {
+                  widget.value = y1
+                }
+                if (widget.name === 'width') {
+                  widget.value = w1
+                }
+                if (widget.name === 'height') {
+                  widget.value = h1
+                }
+              }
+            }
+          }
+
+          this.addWidget('button', 'Set Area', '', () => {
+            let data = {}
+            for (const widget of this.widgets) {
+              if (widget.name === 'x') {
+                data.x = widget.value
+              }
+              if (widget.name === 'y') {
+                data.y = widget.value
+              }
+              if (widget.name === 'width') {
+                data.width = widget.value
+              }
+              if (widget.name === 'height') {
+                data.height = widget.value
+              }
+            }
+            try {
+              console.log('this.inputs', this.id)
+              let imgs = findImages(this.id)
+
+              // let topLinkId = this.inputs[0].link
+              // let topNodeId = app.graph.links[topLinkId].origin_id
+              let topIm = imgs[0]
+
+              let linkId = this.inputs[3].link
+              let nodeId = app.graph.links[linkId].origin_id
+              // console.log(linkId,this.inputs)
+              let imgs2 = findImages(nodeId)
+              let im = imgs2[0]
+              console.log(topIm, im)
+              // let src = im.src
+              setArea(
+                im.naturalWidth,
+                im.naturalHeight,
+                topIm.src,
+                im.src,
+                data,
+                updateValue
+              )
+            } catch (error) {
+              console.log(error)
+            }
+          })
+        }
+      }
+
+      const onRemoved = this.onRemoved
+      this.onRemoved = () => {
+        // let b = this.widgets.filter(w => w.type === 'button')[0];
+
+        return onRemoved?.()
+      }
+
+      if (this.onResize) {
+        this.onResize(this.size)
+      }
+
+      this.serialize_widgets = true //需要保存参数
+    }
+  }
+})
+
+app.registerExtension({
+  name: 'Mixlab.layer.GridInput',
+  async beforeRegisterNodeDef (nodeType, nodeData, app) {
+    if (nodeType.comfyClass == 'GridInput') {
+      const orig_nodeCreated = nodeType.prototype.onNodeCreated
+      nodeType.prototype.onNodeCreated = async function () {
+        orig_nodeCreated?.apply(this, arguments)
+
+        const grids_widget = this.widgets.filter(w => w.name == 'grids')[0]
+
+        const widget = {
+          type: 'div',
+          name: 'upload',
+          draw (ctx, node, widget_width, y, widget_height) {
+            Object.assign(
+              this.div.style,
+              get_position_style(ctx, widget_width, y, node.size[1]),
+              {
+                justifyContent: 'flex-start'
+              }
+            )
+          }
+        }
+
+        widget.div = $el('div', {})
+
+        const addBtn = document.createElement('button')
+        addBtn.innerText = 'Add Box'
+        addBtn.style = `cursor: pointer;
+        font-weight: 300;
+        margin: 2px; 
+        color: var(--descrip-text);
+        background-color: var(--comfy-input-bg);
+        border-radius: 8px;
+        border-color: var(--border-color);
+        border-style: solid;height: 30px;min-width: 122px;
+       `
+
+        const vbtn = document.createElement('button')
+        vbtn.innerText = 'Set Box'
+        vbtn.style = `cursor: pointer;
+        font-weight: 300;
+        margin: 2px; 
+        color: var(--descrip-text);
+        background-color: var(--comfy-input-bg);
+        border-radius: 8px;
+        border-color: var(--border-color);
+        border-style: solid;height: 30px;min-width: 122px;
+       `
+
+        const btn = document.createElement('button')
+        btn.innerText = 'Upload JSON'
+
+        btn.style = `cursor: pointer;
+        font-weight: 300;
+        margin: 2px; 
+        color: var(--descrip-text);
+        background-color: var(--comfy-input-bg);
+        border-radius: 8px;
+        border-color: var(--border-color);
+        border-style: solid;height: 30px;min-width: 122px;
+       `
+
+        addBtn.addEventListener('click', () => {
+          const { width, height, grids } = JSON.parse(grids_widget.value)
+          grids.push({
+            label: 'background',
+            grid: [12, 12, width - 24, height - 24]
+          })
+          grids_widget.value = JSON.stringify(
+            {
+              width,
+              height,
+              grids
+            },
+            null,
+            2
+          )
+        })
+
+        vbtn.addEventListener('click', () => {
+          const { width, height, grids } = JSON.parse(grids_widget.value)
+
+          setAreaTags(width, height, grids, (tag, x, y, w, h) => {
+            grids_widget.value = JSON.stringify(
+              {
+                width,
+                height,
+                grids: Array.from(grids, g => {
+                  if (g.label === tag) {
+                    g.grid = [x, y, w, h]
+                  }
+                  return g
+                })
+              },
+              null,
+              2
+            )
+          })
+        })
+
+        btn.addEventListener('click', () => {
+          let inp = document.createElement('input')
+          inp.type = 'file'
+          inp.accept = '.json'
+          inp.click()
+          inp.addEventListener('change', event => {
+            // 获取选择的文件
+            const file = event.target.files[0]
+            this.title = file.name.split('.')[0]
+
+            // console.log(file.name.split('.')[0])
+            // 创建文件读取器
+            const reader = new FileReader()
+
+            // 定义读取完成事件的回调函数
+            reader.onload = event => {
+              // 读取完成后的文本内容
+              const fileContent = JSON.parse(event.target.result)
+              const grids = fileContent
+              grids_widget.value = JSON.stringify(grids, null, 2)
+              // widget.value = grids
+
+              inp.remove()
+            }
+
+            // 以文本方式读取文件
+            reader.readAsText(file)
+          })
+        })
+
+        widget.div.appendChild(addBtn)
+        widget.div.appendChild(vbtn)
+        widget.div.appendChild(btn)
+        document.body.appendChild(widget.div)
+        this.addCustomWidget(widget)
+
+        const onExecuted = nodeType.prototype.onExecuted
+        nodeType.prototype.onExecuted = function (message) {
+          const r = onExecuted?.apply?.(this, arguments)
+
+          let json = message.json
+          if (json) {
+            json = {
+              width: json[0],
+              height: json[1],
+              grids: json[2]
+            }
+            grids_widget.value = JSON.stringify(json, null, 2)
+            // widget.value = json
+          }
+
+          return r
+        }
+
+        const onRemoved = this.onRemoved
+        this.onRemoved = () => {
+          widget.div.remove()
+          return onRemoved?.()
+        }
+
+        if (this.onResize) {
+          this.onResize(this.size)
+        }
+
+        this.serialize_widgets = true //需要保存参数
+      }
+    }
+  },
+  async loadedGraphNode (node, app) {
+    if (node.type === 'GridInput') {
+      try {
+        const grids_widget = node.widgets.filter(w => w.name == 'grids')[0]
+        const { width, height, grids } = JSON.parse(grids_widget.value)
+        console.log('#GridInput', node, grids)
+
+        const div = node.widgets.filter(w => w.name == 'upload')[0]
+        div.div.querySelector('select').innerHTML = Array.from(
+          grids,
+          g => `<option value="${g.label}">${g.label}</option>`
+        ).join('')
+      } catch (error) {}
+    }
+  }
+})
+
+app.registerExtension({
+  name: 'Mixlab.layer.GridDisplayAndSave',
+  async beforeRegisterNodeDef (nodeType, nodeData, app) {
+    if (nodeType.comfyClass == 'GridDisplayAndSave') {
+      const orig_nodeCreated = nodeType.prototype.onNodeCreated
+      nodeType.prototype.onNodeCreated = async function () {
+        orig_nodeCreated?.apply(this, arguments)
+
+        const grids_widget = this.widgets.filter(w => w.name == 'grids')[0]
+        console.log('GridDisplayAndSave', grids_widget)
+        const widget = {
+          type: 'div',
+          name: 'save_json',
+          draw (ctx, node, widget_width, y, widget_height) {
+            Object.assign(
+              this.div.style,
+              get_position_style(ctx, widget_width, y, node.size[1]),
+              {
+                justifyContent: 'flex-start',
+                flexDirection: 'column'
+              }
+            )
+          }
+        }
+
+        widget.div = $el('div', {})
+
+        const btn = document.createElement('button')
+        btn.innerText = 'Save JSON'
+
+        btn.style = `cursor: pointer;
+        font-weight: 300;
+        margin: 2px; 
+        color: var(--descrip-text);
+        background-color: var(--comfy-input-bg);
+        border-radius: 8px;
+        border-color: var(--border-color);
+        border-style: solid;height: 30px;min-width: 122px;
+        max-width: 122px;
+       `
+
+        btn.addEventListener('click', () => {
+          if (window._mixlab_grid)
+            downloadJsonFile(
+              window._mixlab_grid,
+              this.widgets.filter(w => w.name == 'filename_prefix')[0]?.value +
+                '_grid.json'
+            )
+        })
+
+        widget.div.appendChild(btn)
+        document.body.appendChild(widget.div)
+        this.addCustomWidget(widget)
+
+        const onExecuted = nodeType.prototype.onExecuted
+        nodeType.prototype.onExecuted = function (message) {
+          const r = onExecuted?.apply?.(this, arguments)
+          let save_json = this.widgets.filter(d => d.name == 'save_json')[0]
+          let div = save_json?.div
+          // console.log('Test',message)
+
+          let image = message.image[0]
+          let json = message.json
+          if (image) {
+            const { filename, subfolder, type } = image
+
+            if (!div.querySelector('img')) {
+              let im = new Image()
+              div.appendChild(im)
+              im.style.width = '100%'
+            }
+            div.querySelector('img').src = api.apiURL(
+              `/view?filename=${encodeURIComponent(
+                filename
+              )}&type=${type}&subfolder=${subfolder}${app.getPreviewFormatParam()}${app.getRandParam()}`
+            )
+
+            window._mixlab_grid = {
+              width: json[0],
+              height: json[1],
+              grids: json[2]
+            }
+            // console.log(src)
+          }
+
+          this.onResize?.(this.size)
+
+          return r
+        }
+
+        const onRemoved = this.onRemoved
+        this.onRemoved = () => {
+          widget.div.remove()
+          return onRemoved?.()
+        }
+
+        if (this.onResize) {
+          this.onResize(this.size)
+        }
+
+        this.serialize_widgets = true //需要保存参数
+      }
+    }
+  },
+  async loadedGraphNode (node, app) {
+    if (node.type === 'GridDisplayAndSave') {
+      try {
+        let grids_widget = node.widgets.filter(w => w.name === 'grids')[0]
+        // let ks = getLocalData(`_mixlab_PromptSlide`)
+        let uploadWidget = node.widgets.filter(w => w.name == 'upload')[0]
+        // console.log('##widget', uploadWidget.value)
+        let grids = JSON.parse(uploadWidget.value)
+      } catch (error) {}
+    }
+  }
+})
@@ -3,31 +3,17 @@ import { app } from '../../../scripts/app.js'
 import { ComfyWidgets } from '../../../scripts/widgets.js'
 import { $el } from '../../../scripts/ui.js'

-let api_host = `${window.location.hostname}:${window.location.port}`
-let api_base = ''
-let url = `${window.location.protocol}//${api_host}${api_base}`
+import {
+  getQueue,
+  interrupt,
+  get_position_style,
+  base64Df,
+  getUrl,
+  createImage,
+  sleep
+} from './common.js'

-async function getQueue () {
-  try {
-    const res = await fetch(`${url}/queue`)
-    const data = await res.json()
-    // console.log(data.queue_running,data.queue_pending)
-    return {
-      // Running action uses a different endpoint for cancelling
-      Running: data.queue_running.length,
-      Pending: data.queue_pending.length
-    }
-  } catch (error) {
-    console.error(error)
-    return { Running: 0, Pending: 0 }
-  }
-}
-
-async function interrupt () {
-  const resp = await fetch(`${url}/interrupt`, {
-    method: 'POST'
-  })
-}
+// let url = getUrl()

 async function clipboardWriteImage (win, url) {
  const canvas = document.createElement('canvas')
@@ -208,22 +194,6 @@ async function shareScreen (
  }
 }

-async function sleep (t = 200) {
-  return new Promise((res, rej) => {
-    setTimeout(() => {
-      res(true)
-    }, t)
-  })
-}
-
-function createImage (url) {
-  let im = new Image()
-  return new Promise((res, rej) => {
-    im.onload = () => res(im)
-    im.src = url
-  })
-}
-
 async function compareImages (threshold, previousImage, currentImage) {
  // 将 base64 转换为 Image 对象
  var previousImg = await createImage(previousImage)
@@ -458,44 +428,6 @@ async function requestCamera () {
  return false
 }

-/* 
-A method that returns the required style for the html 
-*/
-function get_position_style (ctx, widget_width, y, node_height) {
-  const MARGIN = 4 // the margin around the html element
-
-  /* Create a transform that deals with all the scrolling and zooming */
-  const elRect = ctx.canvas.getBoundingClientRect()
-  const transform = new DOMMatrix()
-    .scaleSelf(
-      elRect.width / ctx.canvas.width,
-      elRect.height / ctx.canvas.height
-    )
-    .multiplySelf(ctx.getTransform())
-    .translateSelf(MARGIN, MARGIN + y)
-
-  return {
-    transformOrigin: '0 0',
-    transform: transform,
-    left: `0`,
-    top: `0`,
-    cursor: 'pointer',
-    position: 'absolute',
-    maxWidth: `${widget_width - MARGIN * 2}px`,
-    // maxHeight: `${node_height - MARGIN * 2}px`, // we're assuming we have the whole height of the node
-    width: `${widget_width - MARGIN * 2}px`,
-    // height: `${node_height - MARGIN * 2}px`,
-    background: '#EEEEEE',
-    display: 'flex',
-    flexDirection: 'column',
-    // alignItems: 'center',
-    justifyContent: 'space-around'
-  }
-}
-
-const base64Df =
-  'data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAwAAAAMCAYAAABWdVznAAAAAXNSR0IArs4c6QAAALZJREFUKFOFkLERwjAQBPdbgBkInECGaMLUQDsE0AkRVRAYWqAByxldPPOWHwnw4OBGye1p50UDSoA+W2ABLPN7i+C5dyC6R/uiAUXRQCs0bXoNIu4QPQzAxDKxHoALOrZcqtiyR/T6CXw7+3IGHhkYcy6BOR2izwT8LptG8rbMiCRAUb+CQ6WzQVb0SNOi5Z2/nX35DRyb/ENazhpWKoGwrpD6nICp5c2qogc4of+c7QcrhgF4Aa/aoAFHiL+RAAAAAElFTkSuQmCC'
-
 app.registerExtension({
  name: 'Mixlab.image.ScreenShareNode',
  async getCustomWidgets (app) {
@@ -585,9 +517,6 @@ app.registerExtension({
  },
  async beforeRegisterNodeDef (nodeType, nodeData, app) {
    if (nodeType.comfyClass == 'ScreenShare') {
-      /* 
-          Hijack the onNodeCreated call to add our widget
-          */
      const orig_nodeCreated = nodeType.prototype.onNodeCreated
      nodeType.prototype.onNodeCreated = function () {
        orig_nodeCreated?.apply(this, arguments)
@@ -596,22 +525,24 @@ app.registerExtension({
          type: 'HTML', // whatever
          name: 'sreen_share', // whatever
          draw (ctx, node, widget_width, y, widget_height) {
-            console.log('ScreenSHare', y, widget_height)
+            // console.log('ScreenSHare', node)
            Object.assign(
              this.card.style,
-              get_position_style(
-                ctx,
-                widget_width,
-                widget_height * 5,
-                node.size[1]
-              )
+              get_position_style(ctx, widget_width, y, node.size[1], 40)
            )
+ 
          }
        }

-        widget.card = $el('div', {})
+        widget.card = $el('div', {
+          color: 'var(--descrip-text)',
+          backgroundColor: 'var(--comfy-input-bg)'
+        })

-        widget.previewCard = $el('div', {})
+        widget.previewCard = $el('div', {
+          color: 'var(--descrip-text)',
+          backgroundColor: 'var(--comfy-input-bg)'
+        })

        widget.preview = $el('video', {
          style: {
@@ -623,7 +554,10 @@ app.registerExtension({
        })

        widget.previewArea = $el('div', {
-          style: {}
+          style: {
+            color: 'var(--descrip-text)',
+            backgroundColor: 'var(--comfy-input-bg)'
+          }
        })

        widget.shareDiv = $el('div', {
@@ -631,7 +565,9 @@ app.registerExtension({
          style: {
            cursor: 'pointer',
            fontWeight: '300',
-            display: 'flex'
+            display: 'flex',
+            color: 'var(--descrip-text)',
+            backgroundColor: 'var(--comfy-input-bg)'
          }
        })

@@ -642,7 +578,12 @@ app.registerExtension({
            padding: '8px 0',
            fontWeight: '300',
            margin: '2px',
-            width: '100%'
+            width: '100%',
+            color: 'var(--descrip-text)',
+            backgroundColor: 'var(--comfy-input-bg)',
+            borderRadius: '8px',
+            borderColor: 'var(--border-color)',
+            borderStyle: 'solid'
          }
        })

@@ -653,7 +594,12 @@ app.registerExtension({
            padding: '8px 0',
            fontWeight: '300',
            margin: '2px',
-            width: '100%'
+            width: '100%',
+            color: 'var(--descrip-text)',
+            backgroundColor: 'var(--comfy-input-bg)',
+            borderRadius: '8px',
+            borderColor: 'var(--border-color)',
+            borderStyle: 'solid'
          }
        })

@@ -663,29 +609,43 @@ app.registerExtension({
            cursor: 'pointer',
            padding: '8px 0',
            fontWeight: '300',
-            margin: '2px'
+            margin: '2px',
+            color: 'var(--descrip-text)',
+            backgroundColor: 'var(--comfy-input-bg)',
+            borderRadius: '8px',
+            borderColor: 'var(--border-color)',
+            borderStyle: 'solid'
          }
        })

-        widget.refreshInput = $el('input', {
-          placeholder: '  Refresh rate:200 ms',
-          type: 'number',
-          min: 100,
-          step: 100,
-          style: {
-            cursor: 'pointer',
-            padding: '8px 0',
-            fontWeight: '300',
-            margin: '2px'
-          }
-        })
+        // widget.refreshInput = $el('input', {
+        //   placeholder: '  Refresh rate:200 ms',
+        //   type: 'number',
+        //   min: 100,
+        //   step: 100,
+        //   style: {
+        //     cursor: 'pointer',
+        //     padding: '8px 24px',
+        //     fontWeight: '300',
+        //     margin: '2px',
+        //     color: 'var(--descrip-text)',
+        //     backgroundColor: 'var(--comfy-input-bg)'
+        //   }
+        // });
+        // widget.refreshInput.className='comfy-multiline-input'
+
        widget.liveBtn = $el('button', {
          innerText: 'Live Run',
          style: {
            cursor: 'pointer',
            padding: '8px 0',
            fontWeight: '300',
-            margin: '2px'
+            margin: '2px',
+            color: 'var(--descrip-text)',
+            backgroundColor: 'var(--comfy-input-bg)',
+            borderRadius: '8px',
+            borderColor: 'var(--border-color)',
+            borderStyle: 'solid'
          }
        })

@@ -699,7 +659,7 @@ app.registerExtension({
        widget.shareDiv.appendChild(widget.shareBtn)
        widget.shareDiv.appendChild(widget.shareOfWebCamBtn)
        widget.card.appendChild(widget.openFloatingWinBtn)
-        widget.card.appendChild(widget.refreshInput)
+        // widget.card.appendChild(widget.refreshInput)
        widget.card.appendChild(widget.liveBtn)

        const toggleShare = async (isCamera = false) => {
@@ -859,11 +819,11 @@ app.registerExtension({
          toggleShare()
        })

-        widget.refreshInput.addEventListener('change', async () => {
-          window._mixlab_screen_refresh_rate = Math.round(
-            widget.refreshInput.value
-          )
-        })
+        // widget.refreshInput.addEventListener('change', async () => {
+        //   window._mixlab_screen_refresh_rate = Math.round(
+        //     widget.refreshInput.value
+        //   )
+        // })

        widget.liveBtn.addEventListener('click', async () => {
          if (window._mixlab_stopLive) {
@@ -890,6 +850,9 @@ app.registerExtension({
        })
        // console.log('widget.inputEl',widget.inputEl)

+        this.setSize([this.size[0], this.size[1] + 450])
+        app.canvas.draw(true, true)
+
        /*
              Add the widget, make sure we clean up nicely, and we do not want to be serialized!
              */
@@ -901,12 +864,21 @@ app.registerExtension({
          widget.shareBtn.remove()
          widget.liveBtn.remove()
          widget.card.remove()
-          widget.refreshInput.remove()
+          // widget.refreshInput.remove()
          widget.previewArea.remove()
          widget.previewCard.remove()
        }
        this.serialize_widgets = true
      }
+
+      const onExecuted = nodeType.prototype.onExecuted
+      nodeType.prototype.onExecuted = function (message) {
+        onExecuted?.apply(this, arguments)
+        // console.log('###ScreenShare', this, message.refresh_rate)
+        window._mixlab_screen_refresh_rate = Math.round(
+          message.refresh_rate[0] || 500
+        )
+      }
    }
  }
 })
@@ -998,11 +970,13 @@ async function setArea (src) {
  div.innerHTML = `
    <div id='ml_overlay' style='position: absolute;top:0;background: #251f1fc4;
    height: 100vh;
+    z-index:99999999999999;
    width: 100%;'>
      <img id='ml_video' style='position: absolute; 
      height: ${displayHeight}px;user-select: none; 
      -webkit-user-drag: none;
      outline: 2px solid #eaeaea;
+      left: 0;
      box-shadow: 8px 9px 17px #575757;' />
      <div id='ml_selection' style='position: absolute; 
      border: 2px dashed red; 
@@ -1029,12 +1003,7 @@ async function setArea (src) {
    height = displayHeight
  let imgWidth = im.naturalWidth
  let imgHeight = im.naturalHeight
-  // console.log(
-  //   '#screen_share::使用上一次选区 selection',
-  //   data,
-  //   imgWidth,
-  //   img.width
-  // )
+
  if (
    data &&
    data.width > 0 &&
@@ -1048,9 +1017,6 @@ async function setArea (src) {
    y = (img.height * data.y) / data.imgHeight
    width = (img.width * data.width) / data.imgWidth
    height = (img.height * data.height) / data.imgHeight
-    // imgWidth = data.imgWidth
-    // imgHeight = data.imgHeight;
-    // console.log('#screen_share::使用上一次选区 selection', x, y, width, height)
  }

  selection.style.left = x + 'px'
@@ -1115,9 +1081,6 @@ async function setArea (src) {
    let realEndX = (endX / img.offsetWidth) * imgWidth
    let realEndY = (endY / img.offsetHeight) * imgHeight

-    // 输出结果到控制台
-    // console.log('真实宽度: ' + realWidth)
-    // console.log('真实高度: ' + realHeight)
    startX = realStartX
    startY = realStartY
    endX = realEndX
@@ -1127,19 +1090,6 @@ async function setArea (src) {
    let height = Math.abs(endY - startY)
    let left = Math.min(startX, endX)
    let top = Math.min(startY, endY)
-    // Output results to console
-    // console.log('坐标位置: (' + left + ', ' + top + ')')
-    // console.log('宽度: ' + width)
-    // console.log('高度: ' + height)
-
-    // img.removeEventListener('mousedown', startSelection)
-    // img.removeEventListener('mousemove', updateSelection)
-    // img.removeEventListener('mouseup', endSelection)
-
-    // window._mixlab_screen_x = left
-    // window._mixlab_screen_y = top
-    // window._mixlab_screen_width = width
-    // window._mixlab_screen_height = height

    if (width <= 0 && height <= 0) return remove()

@@ -1151,7 +1101,6 @@ async function setArea (src) {
      window._mixlab_screen_webcamVideo,
      !window._mixlab_screen_live
    )
-
    remove()
  }
 }
@@ -1198,10 +1147,10 @@ app.registerExtension({
          type: 'video',
          name: 'FloatingVideo',
          draw (ctx, node, widget_width, y, widget_height) {
-            Object.assign(
-              this.card.style,
-              get_position_style(ctx, widget_width, y, node.size[1])
-            )
+            Object.assign(this.card.style, {
+              ...get_position_style(ctx, widget_width, y, node.size[1], 40),
+              top: `${widget_height}px`
+            })
          }
        }

@@ -1246,13 +1195,18 @@ app.registerExtension({
        })

        widget.PictureInPicture = $el('button', {
-          innerText: 'PictureInPicture',
+          innerText: 'Picture In Picture',
          style: {
            display: 'pictureInPictureEnabled' in document ? 'block' : 'none',
            cursor: 'pointer',
            padding: '8px 0',
            fontWeight: '300',
-            margin: '2px'
+            margin: '2px',
+            color: 'var(--descrip-text)',
+            backgroundColor: 'var(--comfy-input-bg)',
+            borderRadius: '8px',
+            borderColor: 'var(--border-color)',
+            borderStyle: 'solid'
          }
        })

@@ -1305,7 +1259,13 @@ app.registerExtension({
          let w = 360,
            s = widget.preview.videoWidth / widget.preview.videoHeight,
            h = w / s || w
-          console.log(h)
+          // console.log(h)
+
+          if (!window.documentPictureInPicture) {
+            window.alert(
+              'This feature is available only in secure contexts (HTTPS), in some or all supporting browsers. https://developer.mozilla.org/en-US/docs/Web/API/Document_Picture-in-Picture_API'
+            )
+          }

          const pipWindow = await documentPictureInPicture.requestWindow({
            width: w,
@@ -1418,7 +1378,8 @@ app.registerExtension({
            )

            try {
-              pipWindow.document.querySelector('#info').innerText =window._mixlab_screen_seed_input
+              pipWindow.document.querySelector('#info').innerText =
+                window._mixlab_screen_seed_input
            } catch (error) {
              console.log(error)
            }
@@ -1773,6 +1734,14 @@ const updateUI = node => {
    pw.inputEl.title = `Total of ${prompts.length} prompts`
  } else {
    // 动态添加
+    // console.log('ComfyWidgets',ComfyWidgets.STRING(
+    //   node,
+    //   'prompts',
+    //   ['STRING', { multiline: true }]
+    // ))
+
+    // ComfyWidgets.STRING(this, "", ["", {default:this.properties.text, multiline: true}], app)
+
    const w = ComfyWidgets.STRING(
      node,
      'prompts',
@@ -2058,13 +2027,13 @@ const node = {
  name: 'RandomPrompt',
  async init (app) {
    // Any initial setup to run as soon as the page loads
-    console.log('[logging]', 'extension init')
+    // console.log('[logging]', 'extension init')

    if (window.location.href.match('/?')) {
      const { workflow } = getURLParameters(window.location.href)
      if (workflow)
        get_my_workflow().then(data => {
-          console.log('#get_my_workflow', data)
+          // console.log('#get_my_workflow', data)
          let my_workflow = data.filter(
            d => d.filename == 'my_workflow.json'
          )[0]
@@ -2100,10 +2069,15 @@ const node = {
    // }
  },
  loadedGraphNode (node, app) {
-    // Fires for each node when loading/dragging/etc a workflow json or png
-    // If you break something in the backend and want to patch workflows in the frontend
-    // This is the place to do this
-    // console.log("[logging]", "loaded graph node: ", exportGraph(node.graph));
+    if (node.type === 'RandomPrompt') {
+      try {
+        let max_count = node.widgets.filter(w => w.name === 'max_count')[0]
+        max_count.value = node.widgets_values[0]
+        // console.log('RandomPrompt',max_count,node.widgets_values[0])
+      } catch (error) {
+        console.log(error)
+      }
+    }
  },
  async nodeCreated (node) {
    if (node.type === 'RandomPrompt') {
@@ -2196,7 +2170,7 @@ const node = {
        const r = onExecuted?.apply?.(this, arguments)

        let prompts = message.prompts
-        console.log('executed', message)
+        // console.log('executed', message)
        // console.log('#RandomPrompt', this.widgets)
        const pw = this.widgets.filter(w => w.name === 'prompts')[0]

@@ -2207,7 +2181,7 @@ const node = {
        } else {
          // 动态添加
          const w = ComfyWidgets.STRING(
-            node,
+            this,
            'prompts',
            ['STRING', { multiline: true }],
            app
@@ -0,0 +1,195 @@
+import { app } from '../../../scripts/app.js'
+import { api } from '../../../scripts/api.js'
+import { $el } from '../../../scripts/ui.js'
+
+import { get_position_style } from './common.js'
+
+function base64ToBlobFromURL (base64URL, contentType) {
+  return fetch(base64URL).then(response => response.blob())
+}
+
+async function uploadImage (blob, fileType = '.svg', filename) {
+  // const blob = await (await fetch(src)).blob();
+  const body = new FormData()
+  body.append(
+    'image',
+    new File([blob], (filename || new Date().getTime()) + fileType)
+  )
+
+  const resp = await api.fetchApi('/upload/image', {
+    method: 'POST',
+    body
+  })
+
+  // console.log(resp)
+  let data = await resp.json()
+  let { name, subfolder } = data
+  // let src = api.apiURL(
+  //   `/view?filename=${encodeURIComponent(
+  //     name
+  //   )}&type=input&subfolder=${subfolder}${app.getPreviewFormatParam()}${app.getRandParam()}`
+  // )
+
+  return data
+}
+// 上传得到url
+async function uploadBase64ToFile (base64) {
+  let bg_blob = await base64ToBlobFromURL(base64)
+  let url = await uploadImage(bg_blob, '.png')
+  return url
+}
+
+const p5InputNode = {
+  name: 'Mixlab.Comfy.P5Input',
+  async getCustomWidgets (app) {
+    return {
+      IMAGEBASE64 (node, inputName, inputData, app) {
+        const widget = {
+          value: {
+            images: []
+          }, // 不能[x,x,x]
+          type: inputData[0], // the type
+          name: inputName, // the name, slice
+          size: [320, 120], // a default size
+          draw (ctx, node, width, y) {},
+          computeSize (...args) {
+            return [128, 32] // a method to compute the current size of the widget
+          }
+        }
+        node.addCustomWidget(widget)
+        return widget
+      }
+    }
+  },
+
+  async beforeRegisterNodeDef (nodeType, nodeData, app) {
+    if (nodeType.comfyClass == 'P5Input') {
+      // console.log('P5Input')
+      const orig_nodeCreated = nodeType.prototype.onNodeCreated
+
+      nodeType.prototype.onNodeCreated = function () {
+        orig_nodeCreated?.apply(this, arguments)
+
+        const widget = {
+          type: 'div',
+          name: 'image_base64',
+          draw (ctx, node, widget_width, y, widget_height) {
+            Object.assign(
+              this.div.style,
+              get_position_style(
+                ctx,
+                widget_width - 24,
+                44,
+                node.size[1] * 2.8,
+                44
+              )
+            )
+          },
+          serialize: false
+        }
+
+        widget.div = $el('div', {})
+
+        widget.div.style = `margin:12px;width:400px;height:480px;background:white`
+
+        document.body.appendChild(widget.div)
+
+        this.addCustomWidget(widget)
+
+        // document.addEventListener('wheel', handleMouseWheel)
+
+        const onRemoved = this.onRemoved
+        this.onRemoved = () => {
+          widget.div.remove()
+          // window.removeEventListener('message', ms)
+          return onRemoved?.()
+        }
+
+        // 节点的大小控制
+        this.setSize([480, 560])
+        app.canvas.draw(true, true)
+
+        const onResize = this.onResize
+        this.onResize = () => {
+          // 设置最小尺寸
+          if (
+            Math.max(this.size[0], 480) != this.size[0] &&
+            Math.max(this.size[1], 560) != this.size[1]
+          ) {
+            this.setSize([
+              Math.max(this.size[0], 480),
+              Math.max(this.size[1], 560)
+            ])
+          }
+
+          return onResize?.apply(this, arguments)
+        }
+
+        this.serialize_widgets = true //需要保存参数
+      }
+
+      const onExecuted = nodeType.prototype.onExecuted
+      nodeType.prototype.onExecuted = function (message) {
+        onExecuted?.apply(this, arguments)
+        // console.log('##onExecuted', this, message._info)
+        //   app.graph.getNodeById(8).widgets[1].div.querySelector('iframe').contentWindow.postMessage('Hello from parent', '*');
+      }
+    }
+  },
+
+  async loadedGraphNode (node, app) {
+    if (node.type === 'P5Input') {
+    }
+  },
+  nodeCreated (node, app) {
+    //数据延迟？？
+    setTimeout(() => {
+      let widget = node.widgets?.filter(w => w.name == 'image_base64')[0]
+      let framesWidget = node.widgets?.filter(w => w.name == 'frames')[0]
+      if (node.type === 'P5Input' && widget) {
+        console.log('#nodeCreated P5Input')
+        if (framesWidget && !framesWidget.value)
+          framesWidget.value = { images: [] }
+
+        framesWidget.value._seed = Math.random()
+
+        let nodeId = node.id
+        //延迟才能获得this.id
+        widget.div.innerHTML = `<iframe src="mixlab/app/p5_export/p5.html?id=${nodeId}" 
+          style="border:0;width:100%;height:100%;"
+         ></iframe>`
+
+        // 监听来自iframe的消息
+        const ms = async event => {
+          const data = event.data
+          console.log('#P5 Input #', data)
+          if (
+            data.from === 'p5.widget' &&
+            data.status === 'save' &&
+            data.frames &&
+            data.frames.length >= 0 &&
+            data.nodeId == nodeId &&
+            data.id != framesWidget.value.id
+          ) {
+            const frames = data.frames
+
+            //workflow会存储到local，会卡死
+            framesWidget.value.images = []
+            for (const f of frames) {
+              let file = await uploadBase64ToFile(f)
+              framesWidget.value.images.push(file)
+            }
+            // framesWidget.value.base64 = frames
+            // framesWidget.value._seed = Math.random()
+            node.title = 'P5 Input #' + frames.length
+            framesWidget.value.id = data.id
+          }
+        }
+
+        window.addEventListener('message', ms)
+      }
+    }, 1000)
+  }
+}
+
+app.registerExtension(p5InputNode)
--- a/Show More
+++ b/Show More