yresearch/swd-sdxl

Scale-wise Distillation SDXL

Scale-wise Distillation (SwD) is a novel framework for accelerating diffusion models (DMs) by progressively increasing spatial resolution during the generation process.
SwD achieves significant speedups (2.5× to 10×) compared to full-resolution models while maintaining or even 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

Usage

Upgrade to the latest version of the 🧨 diffusers and 🧨 peft

pip install -U diffusers
pip install -U peft

and then you can run

import torch
from diffusers import DDPMScheduler, StableDiffusionXLPipeline
from peft import PeftModel

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    custom_pipeline="quickjkee/swd_pipeline_sdxl",
).to("cuda")

pipe.scheduler = DDPMScheduler.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    subfolder="scheduler",
)

lora_path = "yresearch/swd-sdxl"
pipe.unet = PeftModel.from_pretrained(
    pipe.unet,
    lora_path,
)

prompt = "Cute winter dragon baby, kawaii, Pixar, ultra detailed, glacial background, extremely realistic."
sigmas = [1.0000, 0.8000, 0.6000, 0.4000, 0.0000]
scales = [64, 80, 96, 128]

image = pipe(
    prompt,
    timesteps=torch.tensor(sigmas) * 1000,
    scales=scales,
    height=1024,
    width=1024,
    guidance_scale=1.0,
).images[0]

Citation

@inproceedings{
    starodubcev2026scalewise,
    title={Scale-wise Distillation of Diffusion Models},
    author={Nikita Starodubcev and Ilya Drobyshevskiy and Denis Kuznedelev and Artem Babenko and Dmitry Baranchuk},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026},
    url={https://openreview.net/forum?id=Z06LNjqU1g}
}