From a0ba11bc0e93a3cad0c3644a6dcbedb143c2d652 Mon Sep 17 00:00:00 2001 From: Bubbliiiing <47347516+bubbliiiing@users.noreply.github.com> Date: Mon, 12 Jan 2026 15:06:11 +0800 Subject: [PATCH] Update Qwen-Image-2512 Control (#426) --- README.md | 18 +- README_ja-JP.md | 19 +- README_zh-CN.md | 18 +- config/qwenimage/qwenimage_control.yaml | 5 + examples/qwenimage_fun/predict_i2i_inpaint.py | 243 ++++++ examples/qwenimage_fun/predict_t2i_control.py | 243 ++++++ videox_fun/models/__init__.py | 1 + .../models/qwenimage_transformer2d_control.py | 288 ++++++ videox_fun/pipeline/__init__.py | 1 + .../pipeline/pipeline_qwenimage_control.py | 822 ++++++++++++++++++ 10 files changed, 1640 insertions(+), 18 deletions(-) create mode 100644 config/qwenimage/qwenimage_control.yaml create mode 100644 examples/qwenimage_fun/predict_i2i_inpaint.py create mode 100644 examples/qwenimage_fun/predict_t2i_control.py create mode 100644 videox_fun/models/qwenimage_transformer2d_control.py create mode 100644 videox_fun/pipeline/pipeline_qwenimage_control.py diff --git a/README.md b/README.md index 9bf7769..1a1a9cb 100755 --- a/README.md +++ b/README.md @@ -611,40 +611,46 @@ V1.0: | Qwen-Image-Edit | [🤗Link](https://huggingface.co/Qwen/Qwen-Image-Edit) | [😄Link](https://modelscope.cn/models/Qwen/Qwen-Image-Edit) | Official Qwen-Image-Edit weights | | Qwen-Image-Edit-2509 | [🤗Link](https://huggingface.co/Qwen/Qwen-Image-Edit-2509) | [😄Link](https://modelscope.cn/models/Qwen/Qwen-Image-Edit-2509) | Official Qwen-Image-Edit-2509 weights | -## 7. Z-Image +## 7. Qwen-Image-Fun + +| Name | Storage | Hugging Face | Model Scope | Description | +|--|--|--|--|--| +| Qwen-Image-2512-Fun-Controlnet-Union | - | [🤗Link](https://huggingface.co/alibaba-pai/Qwen-Image-2512-Fun-Controlnet-Union) | [😄Link](https://modelscope.cn/models/PAI/Qwen-Image-2512-Fun-Controlnet-Union) | ControlNet weights for Qwen-Image-2512, supporting multiple control conditions such as Canny, Depth, Pose, MLSD, Scribble, etc. | + +## 8. Z-Image | Name | Storage | Hugging Face | Model Scope | Description | |--|--|--|--|--| | Z-Image-Turbo | [🤗Link](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo) | [😄Link](https://www.modelscope.cn/models/Tongyi-MAI/Z-Image-Turbo) | Official weights for Z-Image-Turbo | -## 8. Z-Image-Fun +## 9. Z-Image-Fun | Name | Storage | Hugging Face | Model Scope | Description | |--|--|--|--|--| | Z-Image-Turbo-Fun-Controlnet-Union | - | [🤗Link](https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union) | [😄Link](https://modelscope.cn/models/PAI/Z-Image-Turbo-Fun-Controlnet-Union) | ControlNet weights for Z-Image-Turbo, supporting multiple control conditions such as Canny, Depth, Pose, MLSD, etc. | | Z-Image-Turbo-Fun-Controlnet-Union-2.1 | - | [🤗Link](https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.1) | [😄Link](https://modelscope.cn/models/PAI/Z-Image-Turbo-Fun-Controlnet-Union-2.1) | ControlNet weights for Z-Image-Turbo. Compared to the first version, it adds to more layers and has been trained for a longer period. It supports multiple control conditions including Canny, Depth, Pose, MLSD, and more. | -## 9. Flux +## 10. Flux | Name | Storage | Hugging Face | Model Scope | Description | |--|--|--|--|--| | FLUX.1-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.1-dev) | [😄Link](https://www.modelscope.cn/models/black-forest-labs/FLUX.1-dev) | Official FLUX.1-dev weights | | FLUX.2-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.2-dev) | [😄Link](https://www.modelscope.cn/models/black-forest-labs/FLUX.2-dev) | Official FLUX.2-dev weights | -## 10. Flux-Fun +## 11. Flux-Fun | Name | Storage | Hugging Face | Model Scope | Description | |--|--|--|--|--| | Flux.2-dev-Fun-Controlnet-Union | - | [🤗Link](https://huggingface.co/alibaba-pai/FLUX.2-dev-Fun-Controlnet-Union) | [😄Link](https://modelscope.cn/models/PAI/FLUX.2-dev-Fun-Controlnet-Union) | Flux.2-dev control weights, supporting various control conditions such as Canny, Depth, Pose, MLSD, etc. | -## 11. HunyuanVideo +## 12. HunyuanVideo | Name | Storage | Hugging Face | Model Scope | Description | |--|--|--|--|--| | HunyuanVideo | [🤗Link](https://huggingface.co/hunyuanvideo-community/HunyuanVideo) | - | HunyuanVideo-diffusers weights | | HunyuanVideo-I2V | [🤗Link](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-I2V) | - | HunyuanVideo-I2V-diffusers weights | -## 12. CogVideoX-Fun +## 13. CogVideoX-Fun V1.5: diff --git a/README_ja-JP.md b/README_ja-JP.md index a04d889..8a6c6d8 100755 --- a/README_ja-JP.md +++ b/README_ja-JP.md @@ -611,39 +611,46 @@ V1.0: | Qwen-Image-Edit | [🤗Link](https://huggingface.co/Qwen/Qwen-Image-Edit) | [😄Link](https://modelscope.cn/models/Qwen/Qwen-Image-Edit) | Qwen-Image-Edit 公式重み | | Qwen-Image-Edit-2509 | [🤗Link](https://huggingface.co/Qwen/Qwen-Image-Edit-2509) | [😄Link](https://modelscope.cn/models/Qwen/Qwen-Image-Edit-2509) | Qwen-Image-Edit-2509 公式重み | -## 7. Z-Image +## 7. Qwen-Image-Fun + +| 名前 | ストレージ | Hugging Face | Model Scope | 説明 | +|--|--|--|--|--| +| Qwen-Image-2512-Fun-Controlnet-Union | - | [🤗リンク](https://huggingface.co/alibaba-pai/Qwen-Image-2512-Fun-Controlnet-Union) | [😄リンク](https://modelscope.cn/models/PAI/Qwen-Image-2512-Fun-Controlnet-Union) | Qwen-Image-2512のControlNet重み。Canny、Depth、Pose、MLSD、Scribbleなど、複数の制御条件をサポートします。 | + +## 8. Z-Image | 名称 | ストレージ | Hugging Face | Model Scope | 説明 | |--|--|--|--|--| | Z-Image-Turbo | [🤗リンク](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo) | [😄リンク](https://www.modelscope.cn/models/Tongyi-MAI/Z-Image-Turbo) | Z-Image-Turboの公式重み | -## 8. Z-Image-Fun +## 9. Z-Image-Fun | 名称 | ストレージ | Hugging Face | Model Scope | 説明 | |--|--|--|--|--| | Z-Image-Turbo-Fun-Controlnet-Union | - | [🤗リンク](https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union) | [😄リンク](https://modelscope.cn/models/PAI/Z-Image-Turbo-Fun-Controlnet-Union) | Z-Image-Turbo用のControlNet重み。Canny、Depth、Pose、MLSDなど複数の制御条件をサポート。 | | Z-Image-Turbo-Fun-Controlnet-Union-2.1 | - | [🤗リンク](https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.1) | [😄リンク](https://modelscope.cn/models/PAI/Z-Image-Turbo-Fun-Controlnet-Union-2.1) | Z-Image-TurboのControlNet重み。第1版と比較して、より多くの層に追加され、より長時間トレーニングされています。Canny、Depth、Pose、MLSDなど、複数の制御条件をサポートしています。 | -## 9. Flux + +## 10. Flux | 名称 | ストレージ | Hugging Face | Model Scope | 説明 | |--|--|--|--|--| | FLUX.1-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.1-dev) | [😄Link](https://www.modelscope.cn/models/black-forest-labs/FLUX.1-dev)| FLUX.1-dev 公式重み | | FLUX.2-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.2-dev) | [😄Link](https://www.modelscope.cn/models/black-forest-labs/FLUX.2-dev) | FLUX.2-dev 公式重み | -## 10. Flux-Fun +## 11. Flux-Fun | 名前 | ストレージ | Hugging Face | ModelScope | 説明 | |--|--|--|--|--| | Flux.2-dev-Fun-Controlnet-Union | - | [🤗リンク](https://huggingface.co/alibaba-pai/FLUX.2-dev-Fun-Controlnet-Union) | [😄リンク](https://modelscope.cn/models/PAI/FLUX.2-dev-Fun-Controlnet-Union) | Flux.2-dev 用の ControlNet 重みで、Canny、Depth、Pose、MLSD など様々な制御条件をサポートします。 | -## 11. HunyuanVideo +## 12. HunyuanVideo | 名称 | ストレージ | Hugging Face | Model Scope | 説明 | |--|--|--|--|--| | HunyuanVideo | [🤗Link](https://huggingface.co/hunyuanvideo-community/HunyuanVideo) | - | HunyuanVideo-diffusers 公式重み | | HunyuanVideo-I2V | [🤗Link](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-I2V) | - | HunyuanVideo-I2V-diffusers 公式重み | -## 12. CogVideoX-Fun +## 13. CogVideoX-Fun V1.5: diff --git a/README_zh-CN.md b/README_zh-CN.md index 373a5fd..6849704 100755 --- a/README_zh-CN.md +++ b/README_zh-CN.md @@ -600,40 +600,46 @@ V1.0: | Qwen-Image-Edit | [🤗Link](https://huggingface.co/Qwen/Qwen-Image-Edit) | [😄Link](https://modelscope.cn/models/Qwen/Qwen-Image-Edit) | Qwen-Image-Edit官方权重 | | Qwen-Image-Edit-2509 | [🤗Link](https://huggingface.co/Qwen/Qwen-Image-Edit-2509) | [😄Link](https://modelscope.cn/models/Qwen/Qwen-Image-Edit-2509) | Qwen-Image-Edit-2509官方权重 | -## 7. Z-Image +## 7. Qwen-Image-Fun + +| 名称 | 存储 | Hugging Face | Model Scope | 描述 | +|--|--|--|--|--| +| Qwen-Image-2512-Fun-Controlnet-Union | - | [🤗链接](https://huggingface.co/alibaba-pai/Qwen-Image-2512-Fun-Controlnet-Union) | [😄链接](https://modelscope.cn/models/PAI/Qwen-Image-2512-Fun-Controlnet-Union) | Qwen-Image-2512的ControlNet权重,支持多种控制条件,如Canny、Depth、Pose、MLSD、Scribble等。 | + +## 8. Z-Image | 名称 | 存储空间 | Hugging Face | Model Scope | 描述 | |--|--|--|--|--| | Z-Image-Turbo | [🤗Link](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo) | [😄Link](https://www.modelscope.cn/models/Tongyi-MAI/Z-Image-Turbo) | Z-Image-Turbo官方权重 | -## 8. Z-Image-Fun +## 9. Z-Image-Fun | 名称 | 存储空间 | Hugging Face | Model Scope | 描述 | |--|--|--|--|--| | Z-Image-Turbo-Fun-Controlnet-Union | - | [🤗链接](https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union) | [😄链接](https://modelscope.cn/models/PAI/Z-Image-Turbo-Fun-Controlnet-Union) | Z-Image-Turbo 的 ControlNet 权重，支持 Canny、Depth、Pose、MLSD 等多种控制条件。 | | Z-Image-Turbo-Fun-Controlnet-Union-2.1 | - | [🤗链接](https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.1) | [😄链接](https://modelscope.cn/models/PAI/Z-Image-Turbo-Fun-Controlnet-Union-2.1) | Z-Image-Turbo 的 ControlNet 权重，相比第一版在更多层进行添加，也训练了更长时间，支持 Canny、Depth、Pose、MLSD 等多种控制条件。 | -## 9. Flux +## 10. Flux | 名称 | 存储空间 | Hugging Face | Model Scope | 描述 | |--|--|--|--|--| | FLUX.1-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.1-dev) | [😄Link](https://www.modelscope.cn/models/black-forest-labs/FLUX.1-dev) | FLUX.1-dev官方权重 | | FLUX.2-dev | [🤗Link](https://huggingface.co/black-forest-labs/FLUX.2-dev) | [😄Link](https://www.modelscope.cn/models/black-forest-labs/FLUX.2-dev) | FLUX.2-dev官方权重 | -## 10. Flux-Fun +## 11. Flux-Fun | 名称 | 存储 | Hugging Face | 魔搭社区（ModelScope） | 描述 | |--|--|--|--|--| | Flux.2-dev-Fun-Controlnet-Union | - | [🤗链接](https://huggingface.co/alibaba-pai/FLUX.2-dev-Fun-Controlnet-Union) | [😄链接](https://modelscope.cn/models/PAI/FLUX.2-dev-Fun-Controlnet-Union) | Flux.2-dev 的 ControlNet 权重，支持 Canny、Depth、Pose、MLSD 等多种控制条件。 | -## 11. HunyuanVideo +## 12. HunyuanVideo | 名称 | 存储空间 | Hugging Face | Model Scope | 描述 | |--|--|--|--|--| | HunyuanVideo | [🤗Link](https://huggingface.co/hunyuanvideo-community/HunyuanVideo) | - | HunyuanVideo-diffusers权重 | | HunyuanVideo-I2V | [🤗Link](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-I2V) | - | HunyuanVideo-I2V-diffusers权重 | -## 12. CogVideoX-Fun +## 13. CogVideoX-Fun V1.5: diff --git a/config/qwenimage/qwenimage_control.yaml b/config/qwenimage/qwenimage_control.yaml new file mode 100644 index 0000000..22200db --- /dev/null +++ b/config/qwenimage/qwenimage_control.yaml @@ -0,0 +1,5 @@ +format: diffusers +pipeline: qwenimage +transformer_additional_kwargs: + control_layers: [0, 12, 24, 36, 48] + control_in_dim: 132 \ No newline at end of file diff --git a/examples/qwenimage_fun/predict_i2i_inpaint.py b/examples/qwenimage_fun/predict_i2i_inpaint.py new file mode 100644 index 0000000..c4bd2df --- /dev/null +++ b/examples/qwenimage_fun/predict_i2i_inpaint.py @@ -0,0 +1,243 @@ +import os +import sys + +import torch + +from omegaconf import OmegaConf +from diffusers import (FlowMatchEulerDiscreteScheduler) + +current_file_path = os.path.abspath(__file__) +project_roots = [os.path.dirname(current_file_path), os.path.dirname(os.path.dirname(current_file_path)), os.path.dirname(os.path.dirname(os.path.dirname(current_file_path)))] +for project_root in project_roots: + sys.path.insert(0, project_root) if project_root not in sys.path else None + +from videox_fun.dist import set_multi_gpus_devices, shard_model +from videox_fun.models import (AutoencoderKLQwenImage, + Qwen2_5_VLForConditionalGeneration, + Qwen2Tokenizer, QwenImageControlTransformer2DModel) +from videox_fun.pipeline import QwenImageControlPipeline +from videox_fun.utils.fm_solvers import FlowDPMSolverMultistepScheduler +from videox_fun.utils.fm_solvers_unipc import FlowUniPCMultistepScheduler +from videox_fun.utils.fp8_optimization import (convert_model_weight_to_float8, + convert_weight_dtype_wrapper) +from videox_fun.utils.lora_utils import merge_lora, unmerge_lora +from videox_fun.utils.utils import (filter_kwargs, get_image_to_video_latent, get_image_latent, get_image, + get_video_to_video_latent, + save_videos_grid) + +# GPU memory mode, which can be chosen in [model_full_load, model_full_load_and_qfloat8, model_cpu_offload, model_cpu_offload_and_qfloat8, sequential_cpu_offload]. +# model_full_load means that the entire model will be moved to the GPU. +# +# model_full_load_and_qfloat8 means that the entire model will be moved to the GPU, +# and the transformer model has been quantized to float8, which can save more GPU memory. +# +# model_cpu_offload means that the entire model will be moved to the CPU after use, which can save some GPU memory. +# +# model_cpu_offload_and_qfloat8 indicates that the entire model will be moved to the CPU after use, +# and the transformer model has been quantized to float8, which can save more GPU memory. +# +# sequential_cpu_offload means that each layer of the model will be moved to the CPU after use, +# resulting in slower speeds but saving a large amount of GPU memory. +GPU_memory_mode = "model_cpu_offload_and_qfloat8" +# Multi GPUs config +# Please ensure that the product of ulysses_degree and ring_degree equals the number of GPUs used. +# For example, if you are using 8 GPUs, you can set ulysses_degree = 2 and ring_degree = 4. +# If you are using 1 GPU, you can set ulysses_degree = 1 and ring_degree = 1. +ulysses_degree = 1 +ring_degree = 1 +# Use FSDP to save more GPU memory in multi gpus. +fsdp_dit = False +fsdp_text_encoder = False +# Compile will give a speedup in fixed resolution and need a little GPU memory. +# The compile_dit is not compatible with the fsdp_dit and sequential_cpu_offload. +compile_dit = False + +# Config path +config_path = "config/qwenimage/qwenimage_control.yaml" +# Model path +model_name = "models/Diffusion_Transformer/Qwen-Image-2512" + +# Choose the sampler in "Flow", "Flow_Unipc", "Flow_DPM++" +sampler_name = "Flow" + +# Load pretrained model if need +transformer_path = "models/Personalized_Model/Qwen-Image-2512-Fun-Controlnet-Union.safetensors" +vae_path = None +lora_path = None + +# Other params +sample_size = [1728, 992] + +# Use torch.float16 if GPU does not support torch.bfloat16 +# ome graphics cards, such as v100, 2080ti, do not support torch.bfloat16 +weight_dtype = torch.bfloat16 +control_image = "asset/pose.jpg" +inpaint_image = "asset/8.png" +mask_image = "asset/mask.png" +control_context_scale = 0.80 + +# 使用更长的neg prompt如"模糊，突变，变形，失真，画面暗，文本字幕，画面固定，连环画，漫画，线稿，没有主体。"，可以增加稳定性 +# 在neg prompt中添加"安静，固定"等词语可以增加动态性。 +prompt = "画面中央是一位年轻女孩，她拥有一头令人印象深刻的亮紫色长发，发丝在海风中轻盈飘扬，营造出动感而唯美的效果。她的长发两侧各扎着黑色蝴蝶结发饰，增添了几分可爱与俏皮感。女孩身穿一袭纯白色无袖连衣裙，裙摆轻盈飘逸，与她清新的气质完美契合。她的妆容精致自然，淡粉色的唇妆和温柔的眼神流露出恬静优雅的气质。她单手叉腰，姿态自信从容，目光直视镜头，展现出既甜美又不失个性的魅力。背景是一片开阔的海景，湛蓝的海水在阳光照射下波光粼粼，闪烁着钻石般的光芒。天空呈现出清澈的蔚蓝色，点缀着几朵洁白的云朵，营造出晴朗明媚的夏日氛围。画面前景右下角可见粉紫色的小花丛和绿色植物，为整体构图增添了自然生机和色彩层次。整张照片色调明亮清新，紫色头发与白色裙装、蓝色海天形成鲜明而和谐的色彩对比。" +negative_prompt = " " +guidance_scale = 4.0 +seed = 43 +num_inference_steps = 50 +lora_weight = 0.55 +save_path = "samples/qwenimage-t2i-control" + +device = set_multi_gpus_devices(ulysses_degree, ring_degree) +config = OmegaConf.load(config_path) + +transformer = QwenImageControlTransformer2DModel.from_pretrained( + model_name, + subfolder="transformer", + low_cpu_mem_usage=True, + torch_dtype=weight_dtype, + transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']), +).to(weight_dtype) + +if transformer_path is not None: + print(f"From checkpoint: {transformer_path}") + if transformer_path.endswith("safetensors"): + from safetensors.torch import load_file + state_dict = load_file(transformer_path) + else: + state_dict = torch.load(transformer_path, map_location="cpu") + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = transformer.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + +# Get Vae +vae = AutoencoderKLQwenImage.from_pretrained( + model_name, + subfolder="vae" +).to(weight_dtype) + +if vae_path is not None: + print(f"From checkpoint: {vae_path}") + if vae_path.endswith("safetensors"): + from safetensors.torch import load_file + state_dict = load_file(vae_path) + else: + state_dict = torch.load(vae_path, map_location="cpu") + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = vae.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + +# Get tokenizer and text_encoder +tokenizer = Qwen2Tokenizer.from_pretrained( + model_name, subfolder="tokenizer" +) +text_encoder = Qwen2_5_VLForConditionalGeneration.from_pretrained( + model_name, subfolder="text_encoder", torch_dtype=weight_dtype +) + +# Get Scheduler +Chosen_Scheduler = scheduler_dict = { + "Flow": FlowMatchEulerDiscreteScheduler, + "Flow_Unipc": FlowUniPCMultistepScheduler, + "Flow_DPM++": FlowDPMSolverMultistepScheduler, +}[sampler_name] +scheduler = Chosen_Scheduler.from_pretrained( + model_name, + subfolder="scheduler" +) + +pipeline = QwenImageControlPipeline( + vae=vae, + tokenizer=tokenizer, + text_encoder=text_encoder, + transformer=transformer, + scheduler=scheduler, +) + +if ulysses_degree > 1 or ring_degree > 1: + from functools import partial + transformer.enable_multi_gpus_inference() + if fsdp_dit: + shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype) + pipeline.transformer = shard_fn(pipeline.transformer) + print("Add FSDP DIT") + if fsdp_text_encoder: + from functools import partial + from videox_fun.dist import set_multi_gpus_devices, shard_model + shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype, module_to_wrapper=text_encoder.language_model.layers) + text_encoder = shard_fn(text_encoder) + print("Add FSDP TEXT ENCODER") + +if compile_dit: + for i in range(len(pipeline.transformer.transformer_blocks)): + pipeline.transformer.transformer_blocks[i] = torch.compile(pipeline.transformer.transformer_blocks[i]) + print("Add Compile") + +if GPU_memory_mode == "sequential_cpu_offload": + pipeline.enable_sequential_cpu_offload(device=device) +elif GPU_memory_mode == "model_cpu_offload_and_qfloat8": + convert_model_weight_to_float8(transformer, exclude_module_name=["img_in", "txt_in", "timestep"], device=device) + convert_weight_dtype_wrapper(transformer, weight_dtype) + pipeline.enable_model_cpu_offload(device=device) +elif GPU_memory_mode == "model_cpu_offload": + pipeline.enable_model_cpu_offload(device=device) +elif GPU_memory_mode == "model_full_load_and_qfloat8": + convert_model_weight_to_float8(transformer, exclude_module_name=["img_in", "txt_in", "timestep"], device=device) + convert_weight_dtype_wrapper(transformer, weight_dtype) + pipeline.to(device=device) +else: + pipeline.to(device=device) + +generator = torch.Generator(device=device).manual_seed(seed) + +if lora_path is not None: + pipeline = merge_lora(pipeline, lora_path, lora_weight, device=device, dtype=weight_dtype) + +with torch.no_grad(): + if inpaint_image is not None: + inpaint_image_input = get_image_latent(inpaint_image, sample_size=sample_size)[:, :, 0] + else: + inpaint_image_input = torch.zeros([1, 3, sample_size[0], sample_size[1]]) + + if mask_image is not None: + mask_image_input = get_image_latent(mask_image, sample_size=sample_size)[:, :1, 0] + else: + mask_image_input = torch.ones([1, 1, sample_size[0], sample_size[1]]) * 255 + + if control_image is not None: + control_image_input = get_image_latent(control_image, sample_size=sample_size)[:, :, 0] + + sample = pipeline( + prompt, + negative_prompt = negative_prompt, + height = sample_size[0], + width = sample_size[1], + generator = generator, + true_cfg_scale = guidance_scale, + num_inference_steps = num_inference_steps, + + image = inpaint_image_input, + mask_image = mask_image_input, + control_image = control_image_input, + control_context_scale = control_context_scale + ).images + +if lora_path is not None: + pipeline = unmerge_lora(pipeline, lora_path, lora_weight, device=device, dtype=weight_dtype) + +def save_results(): + if not os.path.exists(save_path): + os.makedirs(save_path, exist_ok=True) + + index = len([path for path in os.listdir(save_path)]) + 1 + prefix = str(index).zfill(8) + image_path = os.path.join(save_path, prefix + ".png") + image = sample[0] + image.save(image_path) + +if ulysses_degree * ring_degree > 1: + import torch.distributed as dist + if dist.get_rank() == 0: + save_results() +else: + save_results() \ No newline at end of file diff --git a/examples/qwenimage_fun/predict_t2i_control.py b/examples/qwenimage_fun/predict_t2i_control.py new file mode 100644 index 0000000..69fe07e --- /dev/null +++ b/examples/qwenimage_fun/predict_t2i_control.py @@ -0,0 +1,243 @@ +import os +import sys + +import torch + +from omegaconf import OmegaConf +from diffusers import (FlowMatchEulerDiscreteScheduler) + +current_file_path = os.path.abspath(__file__) +project_roots = [os.path.dirname(current_file_path), os.path.dirname(os.path.dirname(current_file_path)), os.path.dirname(os.path.dirname(os.path.dirname(current_file_path)))] +for project_root in project_roots: + sys.path.insert(0, project_root) if project_root not in sys.path else None + +from videox_fun.dist import set_multi_gpus_devices, shard_model +from videox_fun.models import (AutoencoderKLQwenImage, + Qwen2_5_VLForConditionalGeneration, + Qwen2Tokenizer, QwenImageControlTransformer2DModel) +from videox_fun.pipeline import QwenImageControlPipeline +from videox_fun.utils.fm_solvers import FlowDPMSolverMultistepScheduler +from videox_fun.utils.fm_solvers_unipc import FlowUniPCMultistepScheduler +from videox_fun.utils.fp8_optimization import (convert_model_weight_to_float8, + convert_weight_dtype_wrapper) +from videox_fun.utils.lora_utils import merge_lora, unmerge_lora +from videox_fun.utils.utils import (filter_kwargs, get_image_to_video_latent, get_image_latent, get_image, + get_video_to_video_latent, + save_videos_grid) + +# GPU memory mode, which can be chosen in [model_full_load, model_full_load_and_qfloat8, model_cpu_offload, model_cpu_offload_and_qfloat8, sequential_cpu_offload]. +# model_full_load means that the entire model will be moved to the GPU. +# +# model_full_load_and_qfloat8 means that the entire model will be moved to the GPU, +# and the transformer model has been quantized to float8, which can save more GPU memory. +# +# model_cpu_offload means that the entire model will be moved to the CPU after use, which can save some GPU memory. +# +# model_cpu_offload_and_qfloat8 indicates that the entire model will be moved to the CPU after use, +# and the transformer model has been quantized to float8, which can save more GPU memory. +# +# sequential_cpu_offload means that each layer of the model will be moved to the CPU after use, +# resulting in slower speeds but saving a large amount of GPU memory. +GPU_memory_mode = "model_cpu_offload_and_qfloat8" +# Multi GPUs config +# Please ensure that the product of ulysses_degree and ring_degree equals the number of GPUs used. +# For example, if you are using 8 GPUs, you can set ulysses_degree = 2 and ring_degree = 4. +# If you are using 1 GPU, you can set ulysses_degree = 1 and ring_degree = 1. +ulysses_degree = 1 +ring_degree = 1 +# Use FSDP to save more GPU memory in multi gpus. +fsdp_dit = False +fsdp_text_encoder = False +# Compile will give a speedup in fixed resolution and need a little GPU memory. +# The compile_dit is not compatible with the fsdp_dit and sequential_cpu_offload. +compile_dit = False + +# Config path +config_path = "config/qwenimage/qwenimage_control.yaml" +# Model path +model_name = "models/Diffusion_Transformer/Qwen-Image-2512" + +# Choose the sampler in "Flow", "Flow_Unipc", "Flow_DPM++" +sampler_name = "Flow" + +# Load pretrained model if need +transformer_path = "models/Personalized_Model/Qwen-Image-2512-Fun-Controlnet-Union.safetensors" +vae_path = None +lora_path = None + +# Other params +sample_size = [1728, 992] + +# Use torch.float16 if GPU does not support torch.bfloat16 +# ome graphics cards, such as v100, 2080ti, do not support torch.bfloat16 +weight_dtype = torch.bfloat16 +control_image = "asset/pose.jpg" +inpaint_image = None +mask_image = None +control_context_scale = 0.80 + +# 使用更长的neg prompt如"模糊，突变，变形，失真，画面暗，文本字幕，画面固定，连环画，漫画，线稿，没有主体。"，可以增加稳定性 +# 在neg prompt中添加"安静，固定"等词语可以增加动态性。 +prompt = "画面中央是一位年轻女孩，她拥有一头令人印象深刻的亮紫色长发，发丝在海风中轻盈飘扬，营造出动感而唯美的效果。她的长发两侧各扎着黑色蝴蝶结发饰，增添了几分可爱与俏皮感。女孩身穿一袭纯白色无袖连衣裙，裙摆轻盈飘逸，与她清新的气质完美契合。她的妆容精致自然，淡粉色的唇妆和温柔的眼神流露出恬静优雅的气质。她单手叉腰，姿态自信从容，目光直视镜头，展现出既甜美又不失个性的魅力。背景是一片开阔的海景，湛蓝的海水在阳光照射下波光粼粼，闪烁着钻石般的光芒。天空呈现出清澈的蔚蓝色，点缀着几朵洁白的云朵，营造出晴朗明媚的夏日氛围。画面前景右下角可见粉紫色的小花丛和绿色植物，为整体构图增添了自然生机和色彩层次。整张照片色调明亮清新，紫色头发与白色裙装、蓝色海天形成鲜明而和谐的色彩对比。" +negative_prompt = " " +guidance_scale = 4.0 +seed = 43 +num_inference_steps = 50 +lora_weight = 0.55 +save_path = "samples/qwenimage-t2i-control" + +device = set_multi_gpus_devices(ulysses_degree, ring_degree) +config = OmegaConf.load(config_path) + +transformer = QwenImageControlTransformer2DModel.from_pretrained( + model_name, + subfolder="transformer", + low_cpu_mem_usage=True, + torch_dtype=weight_dtype, + transformer_additional_kwargs=OmegaConf.to_container(config['transformer_additional_kwargs']), +).to(weight_dtype) + +if transformer_path is not None: + print(f"From checkpoint: {transformer_path}") + if transformer_path.endswith("safetensors"): + from safetensors.torch import load_file + state_dict = load_file(transformer_path) + else: + state_dict = torch.load(transformer_path, map_location="cpu") + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = transformer.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + +# Get Vae +vae = AutoencoderKLQwenImage.from_pretrained( + model_name, + subfolder="vae" +).to(weight_dtype) + +if vae_path is not None: + print(f"From checkpoint: {vae_path}") + if vae_path.endswith("safetensors"): + from safetensors.torch import load_file + state_dict = load_file(vae_path) + else: + state_dict = torch.load(vae_path, map_location="cpu") + state_dict = state_dict["state_dict"] if "state_dict" in state_dict else state_dict + + m, u = vae.load_state_dict(state_dict, strict=False) + print(f"missing keys: {len(m)}, unexpected keys: {len(u)}") + +# Get tokenizer and text_encoder +tokenizer = Qwen2Tokenizer.from_pretrained( + model_name, subfolder="tokenizer" +) +text_encoder = Qwen2_5_VLForConditionalGeneration.from_pretrained( + model_name, subfolder="text_encoder", torch_dtype=weight_dtype +) + +# Get Scheduler +Chosen_Scheduler = scheduler_dict = { + "Flow": FlowMatchEulerDiscreteScheduler, + "Flow_Unipc": FlowUniPCMultistepScheduler, + "Flow_DPM++": FlowDPMSolverMultistepScheduler, +}[sampler_name] +scheduler = Chosen_Scheduler.from_pretrained( + model_name, + subfolder="scheduler" +) + +pipeline = QwenImageControlPipeline( + vae=vae, + tokenizer=tokenizer, + text_encoder=text_encoder, + transformer=transformer, + scheduler=scheduler, +) + +if ulysses_degree > 1 or ring_degree > 1: + from functools import partial + transformer.enable_multi_gpus_inference() + if fsdp_dit: + shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype) + pipeline.transformer = shard_fn(pipeline.transformer) + print("Add FSDP DIT") + if fsdp_text_encoder: + from functools import partial + from videox_fun.dist import set_multi_gpus_devices, shard_model + shard_fn = partial(shard_model, device_id=device, param_dtype=weight_dtype, module_to_wrapper=text_encoder.language_model.layers) + text_encoder = shard_fn(text_encoder) + print("Add FSDP TEXT ENCODER") + +if compile_dit: + for i in range(len(pipeline.transformer.transformer_blocks)): + pipeline.transformer.transformer_blocks[i] = torch.compile(pipeline.transformer.transformer_blocks[i]) + print("Add Compile") + +if GPU_memory_mode == "sequential_cpu_offload": + pipeline.enable_sequential_cpu_offload(device=device) +elif GPU_memory_mode == "model_cpu_offload_and_qfloat8": + convert_model_weight_to_float8(transformer, exclude_module_name=["img_in", "txt_in", "timestep"], device=device) + convert_weight_dtype_wrapper(transformer, weight_dtype) + pipeline.enable_model_cpu_offload(device=device) +elif GPU_memory_mode == "model_cpu_offload": + pipeline.enable_model_cpu_offload(device=device) +elif GPU_memory_mode == "model_full_load_and_qfloat8": + convert_model_weight_to_float8(transformer, exclude_module_name=["img_in", "txt_in", "timestep"], device=device) + convert_weight_dtype_wrapper(transformer, weight_dtype) + pipeline.to(device=device) +else: + pipeline.to(device=device) + +generator = torch.Generator(device=device).manual_seed(seed) + +if lora_path is not None: + pipeline = merge_lora(pipeline, lora_path, lora_weight, device=device, dtype=weight_dtype) + +with torch.no_grad(): + if inpaint_image is not None: + inpaint_image_input = get_image_latent(inpaint_image, sample_size=sample_size)[:, :, 0] + else: + inpaint_image_input = torch.zeros([1, 3, sample_size[0], sample_size[1]]) + + if mask_image is not None: + mask_image_input = get_image_latent(mask_image, sample_size=sample_size)[:, :1, 0] + else: + mask_image_input = torch.ones([1, 1, sample_size[0], sample_size[1]]) * 255 + + if control_image is not None: + control_image_input = get_image_latent(control_image, sample_size=sample_size)[:, :, 0] + + sample = pipeline( + prompt, + negative_prompt = negative_prompt, + height = sample_size[0], + width = sample_size[1], + generator = generator, + true_cfg_scale = guidance_scale, + num_inference_steps = num_inference_steps, + + image = inpaint_image_input, + mask_image = mask_image_input, + control_image = control_image_input, + control_context_scale = control_context_scale + ).images + +if lora_path is not None: + pipeline = unmerge_lora(pipeline, lora_path, lora_weight, device=device, dtype=weight_dtype) + +def save_results(): + if not os.path.exists(save_path): + os.makedirs(save_path, exist_ok=True) + + index = len([path for path in os.listdir(save_path)]) + 1 + prefix = str(index).zfill(8) + image_path = os.path.join(save_path, prefix + ".png") + image = sample[0] + image.save(image_path) + +if ulysses_degree * ring_degree > 1: + import torch.distributed as dist + if dist.get_rank() == 0: + save_results() +else: + save_results() \ No newline at end of file diff --git a/videox_fun/models/__init__.py b/videox_fun/models/__init__.py index dd80c57..352f865 100755 --- a/videox_fun/models/__init__.py +++ b/videox_fun/models/__init__.py @@ -32,6 +32,7 @@ from .hunyuanvideo_vae import AutoencoderKLHunyuanVideo from .longcatvideo_transformer3d import LongCatVideoTransformer3DModel from .longcatvideo_vae import AutoencoderKLLongCatVideo from .qwenimage_transformer2d import QwenImageTransformer2DModel +from .qwenimage_transformer2d_control import QwenImageControlTransformer2DModel from .qwenimage_vae import AutoencoderKLQwenImage from .wan_audio_encoder import WanAudioEncoder from .wan_image_encoder import CLIPModel diff --git a/videox_fun/models/qwenimage_transformer2d_control.py b/videox_fun/models/qwenimage_transformer2d_control.py new file mode 100644 index 0000000..6b3d293 --- /dev/null +++ b/videox_fun/models/qwenimage_transformer2d_control.py @@ -0,0 +1,288 @@ +# Modified from https://github.com/ali-vilab/VACE/blob/main/vace/models/wan/wan_vace.py +# -*- coding: utf-8 -*- +# Copyright (c) Alibaba, Inc. and its affiliates. + +from typing import Any, Dict, List, Optional, Tuple + +import torch +import torch.nn as nn +from diffusers.configuration_utils import register_to_config +from diffusers.models.modeling_outputs import Transformer2DModelOutput +from diffusers.utils import (USE_PEFT_BACKEND, is_torch_version, + scale_lora_layers, unscale_lora_layers) + +from .qwenimage_transformer2d import (QwenImageTransformer2DModel, + QwenImageTransformerBlock) + + +class QwenImageControlTransformerBlock(QwenImageTransformerBlock): + def __init__( + self, + dim: int, num_attention_heads: int, attention_head_dim: int, + qk_norm: str = "rms_norm", eps: float = 1e-6, + zero_cond_t: bool = False, block_id=0 + ): + super().__init__(dim, num_attention_heads, attention_head_dim, qk_norm, eps, zero_cond_t) + self.block_id = block_id + if block_id == 0: + self.before_proj = nn.Linear(self.dim, self.dim) + nn.init.zeros_(self.before_proj.weight) + nn.init.zeros_(self.before_proj.bias) + self.after_proj = nn.Linear(self.dim, self.dim) + nn.init.zeros_(self.after_proj.weight) + nn.init.zeros_(self.after_proj.bias) + + def forward(self, c, x, **kwargs): + if self.block_id == 0: + c = self.before_proj(c) + x + all_c = [] + else: + all_c = list(torch.unbind(c)) + c = all_c.pop(-1) + + encoder_hidden_states, c = super().forward(c, **kwargs) + c_skip = self.after_proj(c) + all_c += [c_skip, c] + c = torch.stack(all_c) + return encoder_hidden_states, c + + +class BaseQwenImageTransformerBlock(QwenImageTransformerBlock): + def __init__( + self, + dim: int, num_attention_heads: int, attention_head_dim: int, + qk_norm: str = "rms_norm", eps: float = 1e-6, + zero_cond_t: bool = False, block_id=0 + ): + super().__init__(dim, num_attention_heads, attention_head_dim, qk_norm, eps, zero_cond_t) + self.block_id = block_id + + def forward(self, hidden_states, hints=None, context_scale=1.0, **kwargs): + encoder_hidden_states, hidden_states = super().forward(hidden_states, **kwargs) + if self.block_id is not None: + hidden_states = hidden_states + hints[self.block_id] * context_scale + return encoder_hidden_states, hidden_states + +class QwenImageControlTransformer2DModel(QwenImageTransformer2DModel): + @register_to_config + def __init__( + self, + control_layers=None, + control_in_dim=None, + patch_size: int = 2, + in_channels: int = 64, + out_channels: Optional[int] = 16, + num_layers: int = 60, + attention_head_dim: int = 128, + num_attention_heads: int = 24, + joint_attention_dim: int = 3584, + guidance_embeds: bool = False, # TODO: this should probably be removed + axes_dims_rope: Tuple[int, int, int] = (16, 56, 56), + zero_cond_t: bool = False, + use_additional_t_cond: bool = False, + use_layer3d_rope: bool = False, + ): + super().__init__( + patch_size, in_channels, out_channels, num_layers, attention_head_dim, + num_attention_heads, joint_attention_dim, guidance_embeds, axes_dims_rope, + zero_cond_t, use_additional_t_cond, use_layer3d_rope + ) + + self.control_layers = [i for i in range(0, self.num_layers, 2)] if control_layers is None else control_layers + self.control_in_dim = self.in_dim if control_in_dim is None else control_in_dim + + assert 0 in self.control_layers + self.control_layers_mapping = {i: n for n, i in enumerate(self.control_layers)} + + # blocks + self.transformer_blocks = nn.ModuleList( + [ + BaseQwenImageTransformerBlock( + dim=self.inner_dim, + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + zero_cond_t=zero_cond_t, + block_id=self.control_layers_mapping[i] if i in self.control_layers else None + ) + for i in range(num_layers) + ] + ) + + # control blocks + self.control_blocks = nn.ModuleList( + [ + QwenImageControlTransformerBlock( + dim=self.inner_dim, + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + zero_cond_t=zero_cond_t, + block_id=i + ) + for i in self.control_layers + ] + ) + + # control patch embeddings + self.control_img_in = nn.Linear(self.control_in_dim, self.inner_dim) + + def forward_control( + self, + x, + control_context, + kwargs + ): + # embeddings + c = self.control_img_in(control_context) + + # Context Parallel + if self.sp_world_size > 1: + c = torch.chunk(c, self.sp_world_size, dim=1)[self.sp_world_rank] + + # arguments + new_kwargs = dict(x=x) + new_kwargs.update(kwargs) + + for block in self.control_blocks: + if torch.is_grad_enabled() and self.gradient_checkpointing: + def create_custom_forward(module, **static_kwargs): + def custom_forward(*inputs): + return module(*inputs, **static_kwargs) + return custom_forward + ckpt_kwargs = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + encoder_hidden_states, c = torch.utils.checkpoint.checkpoint( + create_custom_forward(block, **new_kwargs), + c, + **ckpt_kwargs, + ) + else: + encoder_hidden_states, c = block(c, **new_kwargs) + new_kwargs["encoder_hidden_states"] = encoder_hidden_states + + hints = torch.unbind(c)[:-1] + return hints + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor = None, + encoder_hidden_states_mask: torch.Tensor = None, + timestep: torch.LongTensor = None, + img_shapes: Optional[List[Tuple[int, int, int]]] = None, + txt_seq_lens: Optional[List[int]] = None, + guidance: torch.Tensor = None, # TODO: this should probably be removed + attention_kwargs: Optional[Dict[str, Any]] = None, + additional_t_cond=None, + control_context=None, + control_context_scale=1.0, + return_dict: bool = True, + ): + if attention_kwargs is not None: + attention_kwargs = attention_kwargs.copy() + lora_scale = attention_kwargs.pop("scale", 1.0) + else: + lora_scale = 1.0 + + if USE_PEFT_BACKEND: + # weight the lora layers by setting `lora_scale` for each PEFT layer + scale_lora_layers(self, lora_scale) + else: + if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None: + logger.warning( + "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective." + ) + + if isinstance(encoder_hidden_states, list): + encoder_hidden_states = torch.stack(encoder_hidden_states) + encoder_hidden_states_mask = torch.stack(encoder_hidden_states_mask) + + hidden_states = self.img_in(hidden_states) + + timestep = timestep.to(hidden_states.dtype) + + if self.zero_cond_t: + timestep = torch.cat([timestep, timestep * 0], dim=0) + modulate_index = torch.tensor( + [[0] * prod(sample[0]) + [1] * sum([prod(s) for s in sample[1:]]) for sample in img_shapes], + device=timestep.device, + dtype=torch.int, + ) + else: + modulate_index = None + + encoder_hidden_states = self.txt_norm(encoder_hidden_states) + encoder_hidden_states = self.txt_in(encoder_hidden_states) + + if guidance is not None: + guidance = guidance.to(hidden_states.dtype) * 1000 + + temb = ( + self.time_text_embed(timestep, hidden_states, additional_t_cond) + if guidance is None + else self.time_text_embed(timestep, guidance, hidden_states, additional_t_cond) + ) + image_rotary_emb = self.pos_embed(img_shapes, txt_seq_lens, device=hidden_states.device) + + # Context Parallel + if self.sp_world_size > 1: + hidden_states = torch.chunk(hidden_states, self.sp_world_size, dim=1)[self.sp_world_rank] + if image_rotary_emb is not None: + image_rotary_emb = ( + torch.chunk(image_rotary_emb[0], self.sp_world_size, dim=0)[self.sp_world_rank], + image_rotary_emb[1] + ) + + # Arguments + kwargs = dict( + encoder_hidden_states=encoder_hidden_states, + encoder_hidden_states_mask=encoder_hidden_states_mask, + temb=temb, + image_rotary_emb=image_rotary_emb, + joint_attention_kwargs=attention_kwargs, + modulate_index=modulate_index, + ) + hints = self.forward_control( + hidden_states, control_context, kwargs + ) + + for index_block, block in enumerate(self.transformer_blocks): + # Arguments + kwargs = dict( + encoder_hidden_states=encoder_hidden_states, + encoder_hidden_states_mask=encoder_hidden_states_mask, + temb=temb, + image_rotary_emb=image_rotary_emb, + joint_attention_kwargs=attention_kwargs, + modulate_index=modulate_index, + hints=hints, + context_scale=control_context_scale + ) + if torch.is_grad_enabled() and self.gradient_checkpointing: + def create_custom_forward(module, **static_kwargs): + def custom_forward(*inputs): + return module(*inputs, **static_kwargs) + return custom_forward + + ckpt_kwargs = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + + encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(block, **kwargs), + hidden_states, + **ckpt_kwargs, + ) + else: + encoder_hidden_states, hidden_states = block(hidden_states, **kwargs) + + if self.zero_cond_t: + temb = temb.chunk(2, dim=0)[0] + # Use only the image part (hidden_states) from the dual-stream blocks + hidden_states = self.norm_out(hidden_states, temb) + output = self.proj_out(hidden_states) + + if self.sp_world_size > 1: + output = self.all_gather(output, dim=1) + + if USE_PEFT_BACKEND: + # remove `lora_scale` from each PEFT layer + unscale_lora_layers(self, lora_scale) + + return output \ No newline at end of file diff --git a/videox_fun/pipeline/__init__.py b/videox_fun/pipeline/__init__.py index 51ab8b2..fe9dba3 100755 --- a/videox_fun/pipeline/__init__.py +++ b/videox_fun/pipeline/__init__.py @@ -9,6 +9,7 @@ from .pipeline_hunyuanvideo import HunyuanVideoPipeline from .pipeline_hunyuanvideo_i2v import HunyuanVideoI2VPipeline from .pipeline_longcatvideo import LongCatVideoPipeline from .pipeline_qwenimage import QwenImagePipeline +from .pipeline_qwenimage_control import QwenImageControlPipeline from .pipeline_qwenimage_edit import QwenImageEditPipeline from .pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline from .pipeline_wan import WanPipeline diff --git a/videox_fun/pipeline/pipeline_qwenimage_control.py b/videox_fun/pipeline/pipeline_qwenimage_control.py new file mode 100644 index 0000000..4564966 --- /dev/null +++ b/videox_fun/pipeline/pipeline_qwenimage_control.py @@ -0,0 +1,822 @@ +# Modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/qwenimage/pipeline_qwenimage.py +# Copyright 2025 Qwen-Image Team and 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. + +import inspect +import math +from dataclasses import dataclass +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +import torchvision.transforms.functional as TF +from diffusers import FlowMatchEulerDiscreteScheduler +from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback +from diffusers.image_processor import VaeImageProcessor +from diffusers.models.embeddings import get_1d_rotary_pos_embed +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.schedulers import FlowMatchEulerDiscreteScheduler +from diffusers.utils import (BaseOutput, is_torch_xla_available, logging, + replace_example_docstring) +from diffusers.utils.torch_utils import randn_tensor +from diffusers.video_processor import VideoProcessor +from einops import rearrange +from PIL import Image +from transformers import T5Tokenizer + +from ..models import (AutoencoderKLQwenImage, + Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer, + QwenImageControlTransformer2DModel) + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + ``` +""" + +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +@dataclass +class QwenImagePipelineOutput(BaseOutput): + """ + Output class for Stable Diffusion pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] + + +class QwenImageControlPipeline(DiffusionPipeline): + r""" + The QwenImage pipeline for text-to-image generation. + + Args: + transformer ([`QwenImageControlTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen2.5-VL-7B-Instruct`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`QwenTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLQwenImage, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + transformer: QwenImageControlTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae.spatial_compression_ratio, do_normalize=False, do_binarize=True, do_convert_grayscale=True + ) + self.tokenizer_max_length = 1024 + self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + self.prompt_template_encode_start_idx = 34 + self.default_sample_size = 128 + + def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor): + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1) + selected = hidden_states[bool_mask] + split_result = torch.split(selected, valid_lengths.tolist(), dim=0) + + return split_result + + def _get_qwen_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + template = self.prompt_template_encode + drop_idx = self.prompt_template_encode_start_idx + txt = [template.format(e) for e in prompt] + txt_tokens = self.tokenizer( + txt, max_length=self.tokenizer_max_length + drop_idx, padding=True, truncation=True, return_tensors="pt" + ).to(device) + encoder_hidden_states = self.text_encoder( + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask, + output_hidden_states=True, + ) + hidden_states = encoder_hidden_states.hidden_states[-1] + split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.attention_mask) + split_hidden_states = [e[drop_idx:] for e in split_hidden_states] + attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states] + max_seq_len = max([e.size(0) for e in split_hidden_states]) + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states] + ) + encoder_attention_mask = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list] + ) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + return prompt_embeds, encoder_attention_mask + + def encode_prompt( + self, + prompt: Union[str, List[str]], + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + prompt_embeds: Optional[torch.Tensor] = None, + prompt_embeds_mask: Optional[torch.Tensor] = None, + max_sequence_length: int = 1024, + ): + r""" + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, device) + + prompt_embeds = prompt_embeds[:, :max_sequence_length] + prompt_embeds_mask = prompt_embeds_mask[:, :max_sequence_length] + + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + return prompt_embeds, prompt_embeds_mask + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + raise ValueError( + "If `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. Make sure to generate `prompt_embeds_mask` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}") + + @staticmethod + def _pack_latents(latents, batch_size, num_channels_latents, height, width, num_frame=None): + if num_frame is None: + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + else: + latents = latents.view(batch_size, num_channels_latents, num_frame, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 3, 5, 1, 4, 6) + latents = latents.reshape(batch_size, num_frame * (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + def _unpack_latents(latents, height, width, vae_scale_factor, num_frame=None): + batch_size, num_patches, channels = latents.shape + if num_frame is None: + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width) + else: + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, num_frame, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 4, 1, 2, 5, 3, 6) + + latents = latents.reshape(batch_size, channels // (2 * 2), num_frame, height, width) + + + return latents + + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, 1, num_channels_latents, height, width) + + if latents is not None: + return latents.to(device=device, dtype=dtype) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + negative_prompt: Union[str, List[str]] = None, + true_cfg_scale: float = 4.0, + height: Optional[int] = None, + width: Optional[int] = None, + + image: Union[torch.FloatTensor] = None, + mask_image: Union[torch.FloatTensor] = None, + control_image: Union[torch.FloatTensor] = None, + subject_ref_images: Union[torch.FloatTensor] = None, + + num_inference_steps: int = 50, + sigmas: Optional[List[float]] = None, + guidance_scale: float = 1.0, + num_images_per_prompt: int = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.Tensor] = None, + prompt_embeds: Optional[torch.Tensor] = None, + prompt_embeds_mask: Optional[torch.Tensor] = None, + negative_prompt_embeds: Optional[torch.Tensor] = None, + negative_prompt_embeds_mask: Optional[torch.Tensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 512, + control_context_scale: float = 1.0 + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + true_cfg_scale (`float`, *optional*, defaults to 1.0): + When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 3.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + + This parameter in the pipeline is there to support future guidance-distilled models when they come up. + Note that passing `guidance_scale` to the pipeline is ineffective. To enable classifier-free guidance, + please pass `true_cfg_scale` and `negative_prompt` (even an empty negative prompt like " ") should + enable classifier-free guidance computations. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is a list with the generated images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + weight_dtype = self.text_encoder.dtype + + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_prompt_embeds_mask is not None + ) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + prompt_embeds, prompt_embeds_mask = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + if do_true_cfg: + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + latents_mean = (torch.tensor(self.vae.config.latents_mean).view(1, self.vae.config.z_dim, 1, 1, 1)).to(device) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(device) + + # Prepare mask latent variables + if mask_image is not None: + mask_condition = self.mask_processor.preprocess(mask_image, height=height, width=width) + mask_condition = torch.where(mask_condition >= 0.5, + torch.ones_like(mask_condition), + torch.zeros_like(mask_condition)) + mask_condition = torch.tile(mask_condition, [1, 3, 1, 1]).to(dtype=weight_dtype, device=device) + else: + mask_condition = torch.zeros([batch_size, 3, height, width]).to(dtype=weight_dtype, device=device) + + if image is not None: + init_image = self.image_processor.preprocess(image, height=height, width=width) + init_image = init_image.to(dtype=weight_dtype, device=device) * (mask_condition < 0.5) + init_image = init_image.unsqueeze(2) + inpaint_latent = self.vae.encode(init_image)[0].mode() + inpaint_latent = ((inpaint_latent - latents_mean) * latents_std).to(dtype=weight_dtype) + else: + inpaint_latent = torch.zeros((batch_size, num_channels_latents, 1, 2 * (int(height) // (self.vae_scale_factor * 2)), 2 * (int(width) // (self.vae_scale_factor * 2)))).to(device, weight_dtype) + + if control_image is not None: + control_image = self.image_processor.preprocess(control_image, height=height, width=width) + control_image = control_image.to(dtype=weight_dtype, device=device) + control_image = control_image.unsqueeze(2) + control_latents = self.vae.encode(control_image)[0].mode() + control_latents = ((control_latents - latents_mean) * latents_std).to(dtype=weight_dtype) + else: + control_latents = torch.zeros_like(inpaint_latent) + + # Unsqueeze + mask_condition = F.interpolate(1 - mask_condition[:, :1], size=inpaint_latent.size()[-2:], mode='nearest').to(device, weight_dtype) + mask_condition = mask_condition.unsqueeze(2) + + control_context = torch.concat([control_latents, mask_condition, inpaint_latent], dim=1) + control_batch_size, control_num_channels_latents, control_num_length_latents, control_height, control_width = control_context.size() + control_context = self._pack_latents(control_context, control_batch_size, control_num_channels_latents, control_height, control_width, num_frame=control_num_length_latents) + + img_shapes = [ + [ + (1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2) + ], + ] * batch_size + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + if self.attention_kwargs is None: + self._attention_kwargs = {} + + txt_seq_lens = prompt_embeds_mask.sum(dim=1).tolist() if prompt_embeds_mask is not None else None + negative_txt_seq_lens = ( + negative_prompt_embeds_mask.sum(dim=1).tolist() if negative_prompt_embeds_mask is not None else None + ) + + # 6. Denoising loop + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + if do_true_cfg: + latent_model_input = torch.cat([latents] * 2) + prompt_embeds_mask_input = [_negative_prompt_embeds_mask for _negative_prompt_embeds_mask in negative_prompt_embeds_mask] + [_prompt_embeds_mask for _prompt_embeds_mask in prompt_embeds_mask] + prompt_embeds_input = [_negative_prompt_embeds for _negative_prompt_embeds in negative_prompt_embeds] + [_prompt_embeds for _prompt_embeds in prompt_embeds] + img_shapes_input = img_shapes * 2 + txt_seq_lens_input = negative_txt_seq_lens + txt_seq_lens + control_context_input = torch.cat([control_context] * 2) + else: + latent_model_input = latents + prompt_embeds_mask_input = prompt_embeds_mask + prompt_embeds_input = prompt_embeds + img_shapes_input = img_shapes + txt_seq_lens_input = txt_seq_lens + control_context_input = control_context + + if hasattr(self.scheduler, "scale_model_input"): + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latent_model_input.shape[0]) + else: + guidance = None + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]).to(latent_model_input.dtype) + print(latent_model_input.size(), control_context_input.size()) + with torch.cuda.amp.autocast(dtype=latents.dtype), torch.cuda.device(device=latents.device): + noise_pred = self.transformer.forward_bs( + x=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=prompt_embeds_mask_input, + encoder_hidden_states=prompt_embeds_input, + img_shapes=img_shapes_input, + txt_seq_lens=txt_seq_lens_input, + attention_kwargs=self.attention_kwargs, + control_context=control_context_input, + control_context_scale=control_context_scale, + return_dict=False, + ) + + if do_true_cfg: + neg_noise_pred, noise_pred = noise_pred.chunk(2) + comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True) + noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True) + noise_pred = comb_pred * (cond_norm / noise_norm) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents[:, :(height // self.vae_scale_factor // 2 * width // self.vae_scale_factor // 2)], height, width, self.vae_scale_factor, num_frame=1) + latents = latents.to(self.vae.dtype) + latents = latents[:, :, :1] + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + image = self.vae.decode(latents, return_dict=False)[0][:, :, 0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return QwenImagePipelineOutput(images=image) \ No newline at end of file