Qwen-Image LoRA Distillation Acceleration Model

Model Introduction

This model is a distilled and accelerated LoRA version of Qwen-Image. We follow the same training procedure as used in DiffSynth-Studio/Qwen-Image-Distill-Full, but replace the trainable model parameters with LoRA, making it easier to integrate into various image generation frameworks.

The training framework is built on DiffSynth-Studio. The training data consists of 16,000 images generated by the original model using randomly sampled prompts from DiffusionDB. The training process ran for approximately one day on 8 * MI308X GPUs.

Performance Comparison

Inference Code

git clone https://github.com/modelscope/DiffSynth-Studio.git  
cd DiffSynth-Studio
pip install -e .

from diffsynth.pipelines.qwen_image import QwenImagePipeline, ModelConfig
from modelscope import snapshot_download
import torch

snapshot_download("DiffSynth-Studio/Qwen-Image-Distill-LoRA", local_dir="models/DiffSynth-Studio/Qwen-Image-Distill-LoRA")
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/"),
)
pipe.load_lora(pipe.dit, "models/DiffSynth-Studio/Qwen-Image-Distill-LoRA/model.safetensors")

prompt = "Exquisite portrait, underwater girl, flowing blue dress, gently floating hair, translucent lighting, surrounded by bubbles, serene expression, intricate details, dreamy and ethereal."
image = pipe(prompt, seed=0, num_inference_steps=15, cfg_scale=1)
image.save("image.jpg")