Qwen-Image.py

from diffsynth.pipelines.qwen_image import QwenImagePipeline, ModelConfig
from diffsynth import load_state_dict
import torch


pipe = QwenImagePipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[
        ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="transformer/diffusion_pytorch_model*.safetensors"),
        ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="text_encoder/model*.safetensors"),
        ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="vae/diffusion_pytorch_model.safetensors"),
    ],
    tokenizer_config=ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="tokenizer/"),
)
state_dict = load_state_dict("models/train/Qwen-Image_full/epoch-1.safetensors")
pipe.dit.load_state_dict(state_dict)
prompt = "a dog"
image = pipe(prompt, seed=0)
image.save("image.jpg")