Swd Large 6 Steps

🚀 Scale-wise Distillation 3.5 Large

Scale-wise Distillation (SwD) is a novel framework for accelerating diffusion models (DMs). It progressively increases spatial resolution during the generation process, achieving 2.5× to 10× speedups compared to full - resolution models while maintaining or improving image quality.

Project page: https://yandex-research.github.io/swd
GitHub: https://github.com/yandex-research/swd
Demo: https://huggingface.co/spaces/dbaranchuk/Scale-wise-Distillation

🚀 Quick Start

To start using Scale-wise Distillation 3.5 Large, you need to upgrade to the latest version of the 🧨 diffusers library.

pip install -U diffusers

Then you can run the following code. (Probably, you will need to specify the visible device: %env CUDA_VISIBLE_DEVICES=0, for correct loading of LoRAs.)

💻 Usage Examples

Basic Usage

import torch
from diffusers import StableDiffusion3Pipeline
from peft import PeftModel

pipe = StableDiffusion3Pipeline.from_pretrained("stabilityai/stable-diffusion-3.5-large",
                                                torch_dtype=torch.float16,
                                                custom_pipeline='quickjkee/swd_pipeline')
pipe = pipe.to("cuda")
lora_path = 'yresearch/swd-large-6-steps'
pipe.transformer = PeftModel.from_pretrained(
    pipe.transformer,
    lora_path,
)
generator = torch.Generator().manual_seed(10)
prompt = 'A cat holding a sign that reads Sample Faster'
sigmas = [1.0000, 0.9454, 0.8959, 0.7904, 0.7371, 0.6022, 0.0000]
scales = [32, 48, 64, 80, 96, 128]

images = pipe(
    prompt,
    sigmas=torch.tensor(sigmas).to('cuda'),
    timesteps=torch.tensor(sigmas[:-1]).to('cuda') * 1000,
    scales=scales,
    guidance_scale=0.0,
    height=int(scales[0] * 8),
    width=int(scales[0] * 8),
    generator=generator,
).images

📄 License

This project is licensed under the Apache-2.0 license.

📚 Documentation

Citation

@article{starodubcev2025swd,
  title={Scale-wise Distillation of Diffusion Models},
  author={Nikita Starodubcev and Denis Kuznedelev and Artem Babenko and Dmitry Baranchuk},
  journal={arXiv preprint arXiv:2503.16397},
  year={2025}
}