DiffSynth-Studio/Qwen-Image-Distill-Full

1---2license: apache-2.03---4# Qwen-Image Full Distillation Accelerated Model5 6![](./assets/title.jpg)7 8## Model Introduction9 10This model is a distilled and accelerated version of [Qwen-Image](https://www.modelscope.cn/models/Qwen/Qwen-Image). The original model requires 40 inference steps and classifier-free guidance (CFG), resulting in a total of 80 forward passes. In contrast, the distilled accelerated model only requires 15 inference steps without CFG, totaling just 15 forward passes—**achieving approximately 5x speedup**. Of course, the number of inference steps can be further reduced based on requirements, though this may lead to some degradation in generation quality.11 12The training framework is built upon [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio). The training data consists of 16,000 images generated by the original model using randomly sampled prompts from [DiffusionDB](https://www.modelscope.cn/datasets/AI-ModelScope/diffusiondb). The training process was conducted on 8 * MI308X GPUs and took approximately one day.13 14## Performance Comparison15 16||Original Model|Original Model|Accelerated Model|17|-|-|-|-|18|Inference Steps|40|15|15|19|CFG Scale|4|1|1|20|Forward Passes|80|15|15|21|Example 1|![](./assets/image_1_full.jpg)|![](./assets/image_1_original.jpg)|![](./assets/image_1_ours.jpg)|22|Example 2|![](./assets/image_2_full.jpg)|![](./assets/image_2_original.jpg)|![](./assets/image_2_ours.jpg)|23|Example 3|![](./assets/image_3_full.jpg)|![](./assets/image_3_original.jpg)|![](./assets/image_3_ours.jpg)|24 25## Inference Code26 27```shell28git clone https://github.com/modelscope/DiffSynth-Studio.git  29cd DiffSynth-Studio30pip install -e .31```32 33```python34from diffsynth.pipelines.qwen_image import QwenImagePipeline, ModelConfig35import torch36 37 38pipe = QwenImagePipeline.from_pretrained(39    torch_dtype=torch.bfloat16,40    device="cuda",41    model_configs=[42        ModelConfig(model_id="DiffSynth-Studio/Qwen-Image-Distill-Full", origin_file_pattern="diffusion_pytorch_model*.safetensors"),43        ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="text_encoder/model*.safetensors"),44        ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="vae/diffusion_pytorch_model.safetensors"),45    ],46    tokenizer_config=ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="tokenizer/"),47)48prompt = "精致肖像，水下少女，蓝裙飘逸，发丝轻扬，光影透澈，气泡环绕，面容恬静，细节精致，梦幻唯美。"49image = pipe(prompt, seed=0, num_inference_steps=15, cfg_scale=1)50image.save("image.jpg")51```