【Diffusers】Qwen-Image-Edit を使ってみる（FLUX.1 Kontext [dev] との比較）

結果

左が元画像、右が作成画像です。

今回の結果

FLUX.1 Kontext [dev]の結果

こちらの記事を見て下さい。

Pythonスクリプト

import torch
from diffusers import QwenImageEditPipeline
from diffusers.utils import load_image
from diffusers.quantizers import PipelineQuantizationConfig

from decorator import gpu_monitor, time_monitor

@time_monitor
@gpu_monitor(interval=0.5)
def main():
    pipeline_quant_config = PipelineQuantizationConfig(
        quant_backend="bitsandbytes_4bit",
        quant_kwargs={
            "load_in_4bit": True,
            "bnb_4bit_quant_type": "nf4",
            "bnb_4bit_compute_dtype": torch.bfloat16,
            "llm_int8_skip_modules": ["transformer_blocks.0.img_mod"]
        },
        components_to_quantize=["text_encoder", "transformer"]
    )

    pipe = QwenImageEditPipeline.from_pretrained(
        "Qwen/Qwen-Image-Edit",
        quantization_config=pipeline_quant_config,
        torch_dtype=torch.bfloat16
    )

    pipe.enable_model_cpu_offload()

    image = load_image("girl.jpg").convert("RGB")

    prompt = "Make the lady hold a sign that says 'Qwen Image Edit is awesome'"

    image = pipe(
        image,
        prompt,
        num_inference_steps=50
    ).images[0]

    image.save("qwenimage_edit.png")

if __name__=="__main__":
    main()

VRAM使用量と時間の計測はこちらのスクリプトを使いました。

RTX 4090を使っています。

GPU 0 - Used memory: 18.68/23.99 GB
time: 144.84 sec

環境構築

pyproject.tomlを載せておきます。

uvを使うとuv syncだけで環境構築できると思います。

[project]
name = "qwen-image-edit"
version = "0.1.0"
description = "Add your description here"
readme = "README.md"
requires-python = ">=3.12"
dependencies = [
    "accelerate>=1.10.1",
    "bitsandbytes>=0.47.0",
    "diffusers>=0.35.1",
    "nvidia-ml-py>=13.580.65",
    "torch==2.8.0+cu126",
    "torchvision==0.23.0+cu126",
    "transformers>=4.56.1",
]

[[tool.uv.index]]
name = "torch-cuda"
url = "https://download.pytorch.org/whl/cu126"
explicit = true

[tool.uv.sources]
torch = [{ index = "torch-cuda" }]
torchvision = [{ index = "torch-cuda" }]