AdaptiveDiff

Category: Sampler-level acceleration

Overview

AdaptiveDiff aims to skip as many noise prediction steps as possible while keeping the final denoised results identical to the original full-step ones. The skipping strategy is guided by the third-order latent difference that indicates the stability between timesteps during the denoising process, enabling the reuse of previous noise prediction results.

Parameters

Parameter	Type	Default	Range	Description
`sampler_name`	enum	euler	see supported list	Base sampler algorithm
`threshold`	float	0.01	0.0 ~ 1.0	Similarity threshold for skipping steps; lower = higher quality, less acceleration
`max_skip_steps`	int	4	0 ~ 10	Maximum consecutive steps allowed to skip

Parameter Tuning Guide

threshold: Start with 0.01 for high quality. Increase to 0.05~0.1 for more aggressive acceleration (2-3x speedup). Values above 0.2 may introduce visible artifacts.
max_skip_steps: Limits how many consecutive steps can be skipped. Higher values allow more acceleration but risk quality degradation in rapid-change regions. Recommended: 2~5.

Supported Samplers

euler, heun, heunpp2, dpm_2, lms, dpm_fast, dpm_adaptive, dpmpp_2m, ipndm, ipndm_v, deis, res_multistep, gradient_estimation, ddim, euler_ancestral, dpm_2_ancestral

Reference

Paper: Training-Free Adaptive Diffusion with Bounded Difference Approximation Strategy
GitHub: https://github.com/InternScience/AdaptiveDiffusion

Citation

@misc{adaptivediffusion24ye,
  author = {Hancheng Ye and Jiakang Yuan and Renqiu Xia and Xiangchao Yan and Tao Chen and Junchi Yan and Botian Shi and Bo Zhang},
  title = {Training-Free Adaptive Diffusion with Bounded Difference Approximation Strategy},
  year = {2024},
  booktitle = {The Thirty-Eighth Annual Conference on Neural Information Processing Systems}
}

2.1 KiB Raw Permalink Blame History