nodes/fchangjun-Baidu_Meux_ComfyTools
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
Files
03d0f260e73da445abdbb5c02fb9752acfbdb0d4
fchangjun-Baidu_Meux_ComfyT…/README.md
T

17 KiB
Raw Blame History

Baidu Meux ComfyTools

English | 中文

English

Overview

Baidu Meux ComfyTools is a collection of custom nodes that streamline common workflow chores in ComfyUI for the Baidu Meux asset platform.
Current version: 1.5.0

  • MeuxMultiSaveImage: save up to sixteen image batches with optional resizing.
  • MeuxAdvancedImageCrop: crop images by pixels or percentage with optional grid alignment.
  • MeuxSimpleLLMNode: call an external chat-completions style LLM API directly inside a workflow.
  • MeuxImageLoader: drop-in replacement for ComfyUI's Load Image node with optional HTTP/HTTPS downloading and persistence to the input folder.
  • MeuxSmartEmptyLatent: generate a safe-sized empty latent tensor based on target size alignment.
  • MeuxSizePresetSafe: compute safe generation size and return size metadata for downstream nodes.
  • MeuxOutpaintSizePresetSafe: compute safe per-side outpaint expansion aligned to 8/64.
  • MeuxSmartExactResize: smart crop/pad to exact target size with auto mode and padding options.
  • MeuxRMBG (displayed as "Meux RMBG (BiRefNet)"): background removal node based on BiRefNet, outputs RGBA IMAGE and MASK.
  • MeuxRealESRGANUpscale (displayed as “Meux ESRGAN Upscale”): local RealESRGAN upscale node using weights from ComfyUI/models/upscale_models or extra_model_paths.yaml (upscale_models).

The package now uses a modular nodes/ directory so each node is easy to maintain and extend.

Changelog

  • v1.5.0
    • Added MeuxRMBG for local/remote BiRefNet background removal.
    • Added MeuxRealESRGANUpscale for local RealESRGAN image upscaling.
  • v1.4.0
    • Added MeuxOutpaintSizePresetSafe for safe outpaint expansion sizes.
    • Capped size inputs at 4096 for size-related nodes.
  • v1.3.0
    • Added MeuxSmartEmptyLatent, MeuxSizePresetSafe, and MeuxSmartExactResize.
  • v1.2.0
    • Introduced MeuxImageLoader, a drop-in replacement for ComfyUI's Load Image with URL downloading.
  • v1.1.0
    • Restructured package into modular node files under nodes/.
    • Added MeuxAdvancedImageCrop and MeuxSimpleLLMNode registrations to the public export.
    • Updated documentation with usage guides for every node.
  • v1.0.0
    • Initial release with the MeuxMultiSaveImage node.

Installation / Update

  1. Clone or download into your ComfyUI custom node folder:

    cd ComfyUI/custom_nodes/
git clone https://github.com/yourusername/Baidu_Meux_ComfyTools.git

  2. For updates, pull the latest changes:

    cd ComfyUI/custom_nodes/Baidu_Meux_ComfyTools
git pull

  3. Install dependencies:

    cd ComfyUI/custom_nodes/Baidu_Meux_ComfyTools
pip install -r requirements.txt

  4. Restart ComfyUI. The nodes appear under image, image/segmentation, image/upscale, Image/Transform, LLM, and utils/size.

Usage Tutorial

MeuxImageLoader

  1. Set source_type:
    • local: behaves exactly like the stock Load Image node; pick any file under ComfyUI/input.
    • url: paste an HTTP/HTTPS link to download on the fly.
  2. Optional controls:
    • filename_hint: suggest the persisted filename; falls back to a hash of the URL.
    • persist_to_input: keep a copy in ComfyUI/input so other nodes can reference it by name.
    • overwrite_existing: replace an existing file instead of auto-appending numeric suffixes.
    • download_timeout, max_download_mb, verify_ssl: guardrails for network transfers.
  3. Run the node to receive the image tensor plus the derived MASK (alpha or grayscale channel when available), making it a drop-in replacement for ComfyUI's default loader.

MeuxMultiSaveImage

  1. Connect one or more IMAGE tensors to images_1 … images_16. Batches are split automatically.
  2. Set filename_prefix (unsafe characters are cleaned automatically).
  3. Optional toggles:
    • save_individually: keep the sequential counter for each input image.
    • resize_to_same: rescale every image to target_width × target_height using Lanczos.
  4. Run the node. Images are written to the ComfyUI output directory as
    prefix_{slotIndex:03d}_{counter:05d}.png, and UI metadata is returned for gallery preview.

MeuxAdvancedImageCrop

  1. Feed an IMAGE tensor into image.
  2. Choose measurement mode:
    • Pixels: specify absolute crop margins.
    • Percentage: margins are interpreted relative to current width/height.
  3. Optionally snap the crop window by setting align_to to 8 or 16.
  4. Execute to obtain the cropped tensor (with console logging of the crop result).

MeuxSimpleLLMNode

  1. Provide your API key and endpoint (api_url defaults to siliconflow chat completions).
  2. Set the model identifier supported by your provider.
  3. Enter the user_prompt; optionally supply a system_prompt and sampling parameters (temperature, top_p, top_k, penalties).
  4. Run the node. The main output is the assistant message, along with the full JSON payload and token usage.

MeuxSmartEmptyLatent

  1. Set target_width, target_height, and batch_size.
  2. Choose align (8 or 64). The node rounds up to a safe generation size.
  3. Run the node to output an empty LATENT plus gen_width and gen_height.

MeuxSizePresetSafe

  1. Set target_width, target_height, and batch_size.
  2. Choose align (8 or 64). The node rounds up to a safe generation size.
  3. Run the node to output gen_width, gen_height, and the input size metadata.

MeuxOutpaintSizePresetSafe

  1. Set expand_left, expand_right, expand_top, and expand_bottom.
  2. Choose align (8 or 64). Each side rounds up to a safe expansion size.
  3. Run the node to output safe_* values plus the original target_* values.

MeuxSmartExactResize

  1. Feed an IMAGE tensor into image.
  2. Set target_width and target_height.
  3. Choose mode: auto, crop_only, or pad_only.
  4. Optional: anchor for cropping, safe_margin_percent for auto crop safety, and padding_mode for padding style.
  5. Run the node to output a resized image with exact target dimensions.

Meux RMBG (BiRefNet)

  1. Prepare one of these model folders in ComfyUI/models/BiRefNet/ (or configure extra_model_paths.yaml with birefnet/BiRefNet):
    • BiRefNet-portrait (portrait matting)
    • BiRefNet (generic foreground)
  2. Ensure each model directory contains config.json and at least one weight file (*.safetensors or *.bin).
  3. Feed an IMAGE tensor into image, choose model_preset, and set input_size (256-2048).
  4. Optional toggles:
    • apply_mask: multiply RGB by mask before output.
    • invert_mask: invert foreground/background selection.
    • model_override: override preset with absolute path or Hugging Face repo id.
  5. Run the node to get:
    • image: RGBA image (alpha from mask)
    • mask: normalized single-channel mask

Meux ESRGAN Upscale

  1. Download the model weight RealESRGAN_x4plus.pth from the official release:

    wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth

  2. Install it under ComfyUI/models/upscale_models/:

    mkdir -p ComfyUI/models/upscale_models
mv RealESRGAN_x4plus.pth ComfyUI/models/upscale_models/

  3. Feed an IMAGE tensor into image.

  4. Choose scale_mode (2x/3x/4x/6x/8x/custom) and set custom_scale when using custom. Note: the bundled model is a 4x model. Non-4x modes are produced by resizing after 4x inference.

  5. Select model_name from the dropdown (auto-scanned from upscale_models and extra_model_paths.yaml). If you add new models, toggle refresh_model_list once.

  6. Optional: set model_path to an absolute path to override the lookup.

  7. Optional: enable free_gpu_after to release GPU memory after each run (slower but safer for long sessions).

  8. Run the node to output the upscaled image tensor.

Folder Structure

Baidu_Meux_ComfyTools/
├── __init__.py          # Registers all exposed nodes
├── requirements.txt     # Python dependencies
├── rmgb.py              # Optional FastAPI RMBG demo service
└── nodes/
    ├── image_loader.py
    ├── advanced_image_crop.py
    ├── multi_save_image.py
    ├── simple_llm_node.py
    ├── smart_empty_latent.py
    ├── size_preset_safe.py
    ├── outpaint_size_preset_safe.py
    ├── smart_exact_resize.py
    ├── rmbg_birefnet.py
    └── realesrgan_upscale.py

Requirements

  • ComfyUI
  • PyTorch
  • Pillow
  • NumPy
  • Requests (required by MeuxSimpleLLMNode and MeuxImageLoader)
  • transformers (required by MeuxRMBG)
  • torchvision (required by MeuxRMBG)
  • realesrgan (required by MeuxRealESRGANUpscale)

License & Support

Licensed under the MIT License.
Issues and feature requests: GitHub Issues.

中文

概述

Baidu Meux ComfyTools 是一组面向百度 Meux 资产平台、帮助简化 ComfyUI 工作流的自定义节点。
当前版本：1.5.0

  • MeuxMultiSaveImage：一次保存最多 16 组图像，支持可选统一尺寸。
  • MeuxAdvancedImageCrop：按像素或百分比裁剪，可选择 8/16 像素对齐。
  • MeuxSimpleLLMNode：在工作流中调用外部 LLM Chat Completion 接口。
  • MeuxImageLoader：兼容本地/URL 两种来源，可选写入 input 目录，完全替换原生 Load Image 节点。
  • MeuxSmartEmptyLatent：根据目标尺寸对齐规则，生成安全尺寸的空白 latent。
  • MeuxSizePresetSafe：计算安全生成尺寸并输出尺寸元信息，供下游节点使用。
  • MeuxOutpaintSizePresetSafe：计算外扩的安全尺寸（按 8/64 对齐）。
  • MeuxSmartExactResize：智能裁剪/补边到精确尺寸，支持自动模式与多种补边方式。
  • MeuxRMBG（显示为 "Meux RMBG (BiRefNet)"）：基于 BiRefNet 的抠图节点，输出 RGBA IMAGE 与 MASK。
  • MeuxRealESRGANUpscale（显示为“Meux ESRGAN Upscale”）：本地 RealESRGAN 放大节点，默认从 ComfyUI/models/upscale_models 或 extra_model_paths.yaml（upscale_models）读取权重。

项目已改用模块化的 nodes/ 目录，便于后续维护与扩展。

更新日志

  • v1.5.0
    • 新增 MeuxRMBG，支持本地/远端 BiRefNet 抠图。
    • 新增 MeuxRealESRGANUpscale，用于本地 RealESRGAN 放大。
  • v1.4.0
    • 新增 MeuxOutpaintSizePresetSafe，用于安全外扩尺寸计算。
    • 尺寸相关输入上限统一为 4096。
  • v1.3.0
    • 新增 MeuxSmartEmptyLatent、MeuxSizePresetSafe、MeuxSmartExactResize。
  • v1.2.0
    • 新增 MeuxImageLoader，支持 URL 下载、可完全替换原生 Load Image。
  • v1.1.0
    • 重构包结构至 nodes/ 子目录。
    • 注册 MeuxAdvancedImageCrop 与 MeuxSimpleLLMNode 节点。
    • 文档新增全部节点的使用教程。
  • v1.0.0
    • 发布 MeuxMultiSaveImage 节点初版。

安装 / 更新

  1. 克隆或下载到 ComfyUI 自定义节点目录：

    cd ComfyUI/custom_nodes/
git clone https://github.com/yourusername/Baidu_Meux_ComfyTools.git

  2. 更新时执行：

    cd ComfyUI/custom_nodes/Baidu_Meux_ComfyTools
git pull

  3. 安装依赖：

    cd ComfyUI/custom_nodes/Baidu_Meux_ComfyTools
pip install -r requirements.txt

  4. 重启 ComfyUI，节点将出现在 image、image/segmentation、image/upscale、Image/Transform、LLM 与 utils/size 分类下。

使用教程

MeuxImageLoader

  1. 选择 source_type：
    • local：与原生 Load Image 完全一致，从 ComfyUI/input 选择文件。
    • url：粘贴 HTTP/HTTPS 图片地址，节点会在运行时自动下载。
  2. 可选参数：
    • filename_hint：为持久化的文件提供自定义名称，默认使用 URL 哈希。
    • persist_to_input：勾选后会把下载结果写入 ComfyUI/input 以供其他节点复用。
    • overwrite_existing：允许覆盖同名文件，否则自动追加序号。
    • download_timeout、max_download_mb、verify_ssl：用于限制下载时间、大小与验证方式。
  3. 运行后输出图片张量与 MASK（若原图包含 Alpha 或灰度通道），可直接替换 ComfyUI 自带的 Load Image 节点。

MeuxMultiSaveImage

  1. 将一张或多张 IMAGE 张量连接到 images_1 … images_16，批次会自动拆分。
  2. 设置 filename_prefix（系统会自动清理危险字符）。
  3. 可选项：
    • save_individually：为每张图单独编号。
    • resize_to_same：按 target_width × target_height 使用 Lanczos 算法统一尺寸。
  4. 运行后，图像保存到 ComfyUI 输出目录，命名格式
    前缀_{输入序号:03d}_{计数:05d}.png，同时返回 UI 预览信息。

MeuxAdvancedImageCrop

  1. 将 IMAGE 张量接入 image。
  2. 选择 measurement 模式：
    • Pixels：输入像素裁剪边距。
    • Percentage：边距按当前宽/高的百分比计算。
  3. 如需像素对齐，将 align_to 设为 8 或 16。
  4. 运行节点即可得到裁剪后的图像，并在控制台查看裁剪日志。

MeuxSimpleLLMNode

  1. 输入可用的 API Key 和接口地址（默认指向 siliconflow chat completions）。
  2. 设置服务商支持的 model 名称。
  3. 填写 user_prompt；根据需要添加 system_prompt 和采样参数（temperature、top_p、top_k、惩罚项）。
  4. 执行节点将返回：
    • 主输出：模型回复文本；
    • 附加输出：完整 JSON 响应与消耗的 token 数。

MeuxSmartEmptyLatent

  1. 设置 target_width、target_height 与 batch_size。
  2. 选择对齐方式 align（8 或 64），节点会向上取整为安全生成尺寸。
  3. 运行后输出空白 LATENT，以及 gen_width、gen_height。

MeuxSizePresetSafe

  1. 设置 target_width、target_height 与 batch_size。
  2. 选择对齐方式 align（8 或 64），节点会向上取整为安全生成尺寸。
  3. 运行后输出 gen_width、gen_height 以及输入尺寸元信息。

MeuxOutpaintSizePresetSafe

  1. 设置 expand_left、expand_right、expand_top、expand_bottom。
  2. 选择 align（8 或 64），各方向向上取整为安全外扩尺寸。
  3. 运行后输出 safe_* 以及原始 target_* 数值。

MeuxSmartExactResize

  1. 将 IMAGE 张量接入 image。
  2. 设置 target_width 与 target_height。
  3. 选择 mode：auto、crop_only 或 pad_only。
  4. 可选：裁剪锚点 anchor、自动裁剪安全边距 safe_margin_percent、补边方式 padding_mode。
  5. 运行后输出精确尺寸的图像张量。

Meux RMBG (BiRefNet)

  1. 在 ComfyUI/models/BiRefNet/ 下准备模型目录（或在 extra_model_paths.yaml 增加 birefnet/BiRefNet 路径）：
    • BiRefNet-portrait（人像抠图）
    • BiRefNet（通用抠图）
  2. 每个模型目录至少包含 config.json 与一个权重文件（*.safetensors 或 *.bin）。
  3. 将 IMAGE 张量接入 image，选择 model_preset，并设置 input_size（256-2048）。
  4. 可选开关：
    • apply_mask：输出前将 RGB 与 mask 相乘。
    • invert_mask：反转前景/背景。
    • model_override：用绝对路径或 Hugging Face 仓库 ID 覆盖预设。
  5. 运行后输出：
    • image：RGBA 图像（alpha 来自 mask）
    • mask：归一化单通道 mask

Meux ESRGAN Upscale

  1. 从官方 Release 下载 RealESRGAN_x4plus.pth：

    wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth

  2. 安装路径为 ComfyUI/models/upscale_models/：

    mkdir -p ComfyUI/models/upscale_models
mv RealESRGAN_x4plus.pth ComfyUI/models/upscale_models/

  3. 将 IMAGE 张量接入 image。

  4. 选择 scale_mode（2x/3x/4x/6x/8x/custom），使用 custom 时设置 custom_scale。 注意：内置模型为 4x，非 4x 模式是在 4x 推理后再缩放得到。

  5. 通过下拉选择 model_name（自动扫描 upscale_models 与 extra_model_paths.yaml）。若新增模型，勾选一次 refresh_model_list。

  6. 可选：设置 model_path 为绝对路径以覆盖默认查找。

  7. 可选：开启 free_gpu_after，每次运行后释放显存（速度更慢，但适合长时间运行）。

  8. 运行后输出放大后的图像张量。

目录结构

Baidu_Meux_ComfyTools/
├── __init__.py          # 节点入口注册
├── requirements.txt     # Python 依赖
├── rmgb.py              # 可选 FastAPI 抠图演示服务
└── nodes/
    ├── image_loader.py
    ├── advanced_image_crop.py
    ├── multi_save_image.py
    ├── simple_llm_node.py
    ├── smart_empty_latent.py
    ├── size_preset_safe.py
    ├── outpaint_size_preset_safe.py
    ├── smart_exact_resize.py
    ├── rmbg_birefnet.py
    └── realesrgan_upscale.py

依赖

  • ComfyUI
  • PyTorch
  • Pillow
  • NumPy
  • Requests（供 MeuxSimpleLLMNode 与 MeuxImageLoader 访问 HTTP 接口）
  • transformers（供 MeuxRMBG 使用）
  • torchvision（供 MeuxRMBG 使用）
  • realesrgan（供 MeuxRealESRGANUpscale 使用）

许可证与支持

采用 MIT 许可证。
问题反馈与功能建议： GitHub Issues。

Reference in New Issue View Git Blame Copy Permalink