Qwenの画像生成モデル「Qwen−Image」を試す

uv init -p 3.12 qwen-image-work && cd $_

uv add dfloat11[cuda12]

 + dfloat11==0.3.1

uv add git+https://github.com/huggingface/diffusers

 + diffusers==0.35.0.dev0 (from git+https://github.com/huggingface/diffusers@ba2ba9019f76fd96c532240ed07d3f98343e4041)

from diffusers import DiffusionPipeline, QwenImageTransformer2DModel
import torch
from transformers.modeling_utils import no_init_weights
from dfloat11 import DFloat11Model
import argparse

def parse_args():
    parser = argparse.ArgumentParser(description='Generate images using Qwen-Image model')
    parser.add_argument('--cpu_offload', action='store_true', help='Enable CPU offloading')
    parser.add_argument('--no_pin_memory', action='store_true', help='Disable memory pinning')
    parser.add_argument('--prompt', type=str, default='A coffee shop entrance features a chalkboard sign reading "Qwen Coffee 😊 $2 per cup," with a neon light beside it displaying "通义千问". Next to it hangs a poster showing a beautiful Chinese woman, and beneath the poster is written "π≈3.1415926-53589793-23846264-33832795-02384197".',
                        help='Text prompt for image generation')
    parser.add_argument('--negative_prompt', type=str, default=' ',
                        help='Negative prompt for image generation')
    parser.add_argument('--aspect_ratio', type=str, default='16:9', choices=['1:1', '16:9', '9:16', '4:3', '3:4'],
                        help='Aspect ratio of generated image')
    parser.add_argument('--num_inference_steps', type=int, default=50,
                        help='Number of denoising steps')
    parser.add_argument('--true_cfg_scale', type=float, default=4.0,
                        help='Classifier free guidance scale')
    parser.add_argument('--seed', type=int, default=42,
                        help='Random seed for generation')
    parser.add_argument('--output', type=str, default='example.png',
                        help='Output image path')
    parser.add_argument('--language', type=str, default='en', choices=['en', 'zh'],
                        help='Language for positive magic prompt')
    return parser.parse_args()

args = parse_args()

model_name = "Qwen/Qwen-Image"

with no_init_weights():
    transformer = QwenImageTransformer2DModel.from_config(
        QwenImageTransformer2DModel.load_config(
            model_name, subfolder="transformer",
        ),
    ).to(torch.bfloat16)

DFloat11Model.from_pretrained(
    "DFloat11/Qwen-Image-DF11",
    device="cpu",
    cpu_offload=args.cpu_offload,
    pin_memory=not args.no_pin_memory,
    bfloat16_model=transformer,
)

pipe = DiffusionPipeline.from_pretrained(
    model_name,
    transformer=transformer,
    torch_dtype=torch.bfloat16,
)
pipe.enable_model_cpu_offload()

positive_magic = {
    "en": "Ultra HD, 4K, cinematic composition.", # for english prompt,
    "zh": "超清，4K，电影级构图" # for chinese prompt,
}

# Generate with different aspect ratios
aspect_ratios = {
    "1:1": (1328, 1328),
    "16:9": (1664, 928),
    "9:16": (928, 1664),
    "4:3": (1472, 1140),
    "3:4": (1140, 1472),
}

width, height = aspect_ratios[args.aspect_ratio]

image = pipe(
    prompt=args.prompt + positive_magic[args.language],
    negative_prompt=args.negative_prompt,
    width=width,
    height=height,
    num_inference_steps=args.num_inference_steps,
    true_cfg_scale=args.true_cfg_scale,
    generator=torch.Generator(device="cuda").manual_seed(args.seed)
).images[0]

image.save(args.output)

max_memory = torch.cuda.max_memory_allocated()
print(f"Max memory: {max_memory / (1000 ** 3):.2f} GB")

uv run qwen_image.py --help

usage: qwen_image.py [-h] [--cpu_offload] [--no_pin_memory] [--prompt PROMPT] [--negative_prompt NEGATIVE_PROMPT]
                     [--aspect_ratio {1:1,16:9,9:16,4:3,3:4}] [--num_inference_steps NUM_INFERENCE_STEPS]
                     [--true_cfg_scale TRUE_CFG_SCALE] [--seed SEED] [--output OUTPUT] [--language {en,zh}]

Generate images using Qwen-Image model

options:
  -h, --help            show this help message and exit
  --cpu_offload         Enable CPU offloading
  --no_pin_memory       Disable memory pinning
  --prompt PROMPT       Text prompt for image generation
  --negative_prompt NEGATIVE_PROMPT
                        Negative prompt for image generation
  --aspect_ratio {1:1,16:9,9:16,4:3,3:4}
                        Aspect ratio of generated image
  --num_inference_steps NUM_INFERENCE_STEPS
                        Number of denoising steps
  --true_cfg_scale TRUE_CFG_SCALE
                        Classifier free guidance scale
  --seed SEED           Random seed for generation
  --output OUTPUT       Output image path
  --language {en,zh}    Language for positive magic prompt

uv run qwen_image.py \
    --cpu_offload \
    --prompt "Two racehorses are neck and neck as they approach the finish line, with each jockey urging their mount on with everything they've got."

uv run qwen_image.py \
    --cpu_offload \
    --prompt "2頭の競走馬がゴール前で競っている。それぞれの騎手は懸命に馬を追っている。"

uv init -p 3.12 --bare comfyui && cd comfyui

cat <<EOF >> pyproject.toml

[tool.uv.sources]
torch = { index = "pytorch-cu128" }
torchvision = { index = "pytorch-cu128" }
torchaudio = { index = "pytorch-cu128" }

[[tool.uv.index]]
name = "pytorch-cu128"
url = "https://download.pytorch.org/whl/cu128"
explicit = true
EOF

uv add torch torchvision torchaudio

uv add comfy-cli pip

uv run comfy --here install

uv run comfy launch -- --listen 0.0.0.0

comfyui
├── ComfyUI
(snip)
│   ├── models
│   │   ├── checkpoints
│   │   ├── clip
│   │   ├── clip_vision
│   │   ├── configs
│   │   ├── controlnet
│   │   ├── diffusers
│   │   ├── diffusion_models
│   │   │   └── qwen_image_fp8_e4m3fn.safetensors  # ここ
│   │   ├── embeddings
│   │   ├── gligen
│   │   ├── hypernetworks
│   │   ├── loras
│   │   ├── photomaker
│   │   ├── style_models
│   │   ├── text_encoders
│   │   │   └── qwen_2.5_vl_7b_fp8_scaled.safetensors  # ここ
│   │   ├── unet
│   │   ├── upscale_models
│   │   ├── vae
│   │   │   └── qwen_image_vae.safetensors  # ここ
│   │   └── vae_approx
(snip)

comfyui
├── ComfyUI
(snip)
│   ├── output
│   │   └── ComfyUI_00001_.png
(snip)

+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 570.172.08             Driver Version: 570.172.08     CUDA Version: 12.8     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 4090        On  |   00000000:01:00.0 Off |                  Off |
| 57%   76C    P2            430W /  450W |   21172MiB /  24564MiB |    100%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+