SADA: Stability-guided Adaptive Diffusion Acceleration
Authors: Ting Jiang, Yixiao Wang, Hancheng Ye, Zishan Shao, Jingwei Sun, Jingyang Zhang, Zekai Chen, Jianyi Zhang, Yiran Chen, Hai Li
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive evaluations on SD-2, SDXL, and Flux using both EDM and DPM++ solvers reveal consistent 1.8 speedups with minimal fidelity degradation (LPIPS 0.10 and FID 4.5) compared to unmodified baselines, significantly outperforming prior methods. Moreover, SADA adapts seamlessly to other pipelines and modalities: It accelerates Control Net without any modifications and speeds up Music LDM by 1.8 with 0.01 spectrogram LPIPS. Our code is available at: https://github.com/Ting-Justin-Jiang/sadaicml. |
| Researcher Affiliation | Academia | 1Department of Electrical and Computer Engineering, Duke University, Durham, U.S.A 2Department of Statistical Science, Duke University, Durham, U.S.A. Correspondence to: Ting Jiang <EMAIL>. |
| Pseudocode | No | The paper describes methods in sections 3.3, 3.4, and 3.5 but does not include any clearly labeled pseudocode or algorithm blocks. The steps are described in prose. |
| Open Source Code | Yes | Our code is available at: https://github.com/Ting-Justin-Jiang/sadaicml. |
| Open Datasets | Yes | All experiments are conducted using the MSCOCO-2017 validation set as generation prompts under identical conditions to assess efficiency and quality. |
| Dataset Splits | Yes | All experiments are conducted using the MSCOCO-2017 validation set as generation prompts under identical conditions to assess efficiency and quality. |
| Hardware Specification | Yes | Experiments with Flux.1-dev are executed on a single NVIDIA A100 GPU, while the remaining experiments are run on a single NVIDIA A5000 GPU. |
| Software Dependencies | No | All pipelines are implemented using the Huggingface Diffusers framework. While a framework is mentioned, no specific version number for Huggingface Diffusers or other key software components is provided. |
| Experiment Setup | Yes | We perform evaluations using two sampling schedulers Euler Discrete Multistep (EDM) (Karras et al., 2022) Solver (firstorder) and DPM-Solver++ (Lu et al., 2022b) (second-order) each configured with 50 sampling steps. We provide justification for our choice of base step T = 50 in Figure A.3. To evaluate SADA s performance under fewstep sampling, Table 2 reports results on {SD-2, SDXL} using {Euler, DPM-Solver++} while varying the number of inference steps {50, 25, 15}. Note that the Lagrange interpolation parameters are slightly adjusted to match the shorter denoising schedules in these few-step settings. |