SADA

Category: Sampler-level acceleration (pure sampler method, no model patching)

Overview

SADA (Stability-guided Adaptive Diffusion Acceleration) is a novel paradigm that unifies step-wise and token-wise sparsity decisions via a single stability criterion to accelerate sampling of ODE-based generative models (Diffusion and Flow-matching). It adaptively allocates sparsity based on the sampling trajectory, and introduces principled approximation schemes that leverage precise gradient information from the numerical ODE solver.

The original paper proposes both step-level acceleration (adaptive step skipping) and token-level sparsity (token pruning). For generality and compatibility with ComfyUI's unified k-diffusion interface, we only integrate the step-level sampler acceleration — the core momentum-based stability criterion and Lagrange interpolation reconstruction. This allows SADA to work as a pure sampler method with any model type without requiring model patching or architecture-specific modifications.

Parameters

Parameter	Type	Default	Range	Description
`acc_start`	int	10	0 ~ 100	Step index to start acceleration (min 3)
`acc_end`	int	47	-100 ~ 200	Step index to stop acceleration (negative = N+value)
`max_interval`	int	4	1 ~ 20	Maximum consecutive steps to skip
`lagrange_term`	int	3	0 ~ 5	Lagrange interpolation order (0 = disabled)
`lagrange_int`	int	4	1 ~ 20	Lagrange anchor point interval
`lagrange_step`	int	20	5 ~ 100	Step after which Lagrange interpolation activates

Parameter Tuning Guide

acc_start / acc_end: Defines the active acceleration window. Early and final steps are typically not skipped (high-change regions). For 50-step schedules, (10, 47) is a good default.
max_interval: Limits consecutive skips. 3~5 is safe; higher values risk quality drop.
lagrange_term: Set to 3~4 for higher-quality interpolation at skipped steps. 0 disables Lagrange (uses simple x0 prediction).
lagrange_int: Interval between anchor points for Lagrange interpolation. 4~6 is typical.
lagrange_step: Lagrange only activates after this step (early steps need full computation). 20~30 for 50-step schedules. Must satisfy lagrange_step % lagrange_int == 0.

Supported Samplers

euler, heun, heunpp2, dpm_2, lms, dpmpp_2m, ddim, euler_ancestral, dpm_2_ancestral

Reference

Paper: SADA: Stability-guided Adaptive Diffusion Acceleration
GitHub: https://github.com/Ting-Justin-Jiang/sada-icml

Citation

@inproceedings{jiang2025sada,
  title     = {SADA: Stability-guided Adaptive Diffusion Acceleration},
  author    = {Ting Jiang and Yixiao Wang and Hancheng Ye and Zishan Shao and Jingwei Sun and Jingyang Zhang and Zekai Chen and Jianyi Zhang and Yiran Chen and Hai Li},
  booktitle = {Proceedings of the 42nd International Conference on Machine Learning},
  year      = {2025}
}

3.1 KiB Raw Permalink Blame History