Optimal Stepsize for Diffusion Sampling(OSS)

April 13, 2025 · View on GitHub

arXiv 

Official PyTorch implementation for our paper Optimal Stepsize for Diffusion Sampling, a plug-and-play algorithm to search the optimal sampling stepsize in diffusion sampling.

teaser
Left to Right: FLUX-100steps; FLUX+OSS-10steps; FLUX-10steps.

https://github.com/user-attachments/assets/d903c467-7050-4e79-9659-47bb672f001c

Left to Right: Wan-100steps; Wan+OSS-20steps; Wan-20steps.

:fire: Latest News!!

  • Apr 13, 2025: :clap: OSS is integrated into ComfyUI as OptimalStepsScheduler! It supports the fast sampling of text-to-image model FLUX and text-to-video model Wan. Here, we give a comparision of FLUX workflow with 10 sampling steps: pr Here are the demo workflows for FLUX-workflow and Wan-workflow. 10 steps for FLUX and 20 steps for Wan are highly recommended. More details are included in our pull request here.

:smiley: Overview

In this repo, we provide some examples of using our algorithm based on the DiT, FLUX, Open-Sora, and Wan2.1.

Note that OSS is not limited to these, we also provide a guidance to adapt it to other diffusion models.

:rocket: Quick Start Guide

Step 0: Prepare the environment

Prepare the environment for the target model you want to use, such as DiT, FLUX, Open-Sora, Wan2.1 or other models.

Step 1: Clone the repository

git clone https://github.com/bebebe666/OSS.git
cd OSS

Step 2: Run inference

# DiT 
bash scripts/dit.sh
# FLUX
bash scripts/flux.sh
# Open-Sora
bash scripts/opensora.sh
# Wan2.1
bash scripts/wan.sh

:airplane: Appling for other models

Model-Preparing

Before using our algorithm, you need to wrap your diffusion model to a unified format, this should satisfies:

  • The output of the model should be the v-pred same as the Flow Matching.
  • The sampling trajectory should be straight, follows xj=xi+v(tj−ti)x_j = x_i + v(t_j - t_i).

You can refer to examples we gave in the model_wrap.py.

Searching

We provide the searching functions as follows:

oss_steps = search_OSS(model, z, search_batch, context, device, teacher_steps=10, student_steps=5, model_kwargs=model_kwargs)
  • the z is the input noise;
  • the search_batch is the number of images you want to search;
  • the context is the class embedding in class conditional image generation and prompt embedding in text-to-image generation.

We provide the function search_OSS_video for video generation model searching, which supports the selection of frame and channel. As default, cost_type="all" and channel_type="all" means using all the frames and channels for cost calculation. You can pass any number (remember not greater than the total) as a string format like "4" to use part of them.

Inference

After getting the oss_steps, you can pass it to the inference function to get the sampling results.

samples = infer_OSS(oss_steps, model, z, context, device, model_kwargs=model_kwargs)

🙏 Acknowledgments

This codebase benefits from the solid prior works: DiT, FLUX, Open-Sora, and Wan2.1 for their excellent generation ability.

📖 Citation

If you find this project helpful for your research or use it in your own work, please cite our paper:

@misc{pei2025optimalstepsizediffusionsampling,
      title={Optimal Stepsize for Diffusion Sampling}, 
      author={Jianning Pei and Han Hu and Shuyang Gu},
      year={2025},
      eprint={2503.21774},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.21774}, 
}

⭐️ If this repository helped your research, please star 🌟 this repo 👍!