Research map / Image generation / Fast diffusion sampling
SADA: Stability-guided Adaptive Diffusion Acceleration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Training-free SADA accelerates ODE-based diffusion and flow-matching sampling by unifying step-wise and token-wise sparsity under a second-order-difference stability criterion, with Adams-Moulton and Lagrange-interpolation approximations.
Problem
Training-free acceleration methods that reduce per-step or step count cost in diffusion models lose fidelity relative to the unmodified baseline, because fixed sparsity patterns ignore prompt-dependent trajectories and the underlying ODE solver.
Approach
Uses a stability criterion built from the second-order difference of the trajectory velocity and the third-order extrapolation error to decide, per timestep, between step-wise caching, multistep interpolation, and token-wise pruning, with approximations derived from the ODE solver (Adams-Moulton, Lagrange interpolation).
Key results
- On SD-2 with DPM++, SADA reaches 1.80x speedup with LPIPS 0.094 and FID 4.02, versus AdaptiveDiffusion at 1.45x with LPIPS 0.100 and FID 4.35.
- On SDXL with DPM++, SADA reduces FID from 8.48 (DeepCache) to 3.51 at 1.86x speedup.
- On Flux.1-dev, SADA attains LPIPS 0.060 and FID 1.95 at 2.02x speedup, versus TeaCache's LPIPS 0.216 and FID 4.89 at 2.00x.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SADA instead of AdaptiveDiffusion | SD-2 with DPM++ · MS-COCO 2017 | FID: 4.35 -> 4.02 | Helped |
| use SADA instead of AdaptiveDiffusion | SD-2 with Euler · MS-COCO 2017 | FID: 7.58 -> 4.26 | Helped |
| use SADA instead of AdaptiveDiffusion | SDXL with DPM++ · MS-COCO 2017 | FID: 4.59 -> 3.51 | Helped |
| use SADA instead of AdaptiveDiffusion | SDXL with Euler · MS-COCO 2017 | FID: 6.11 -> 3.76 | Helped |
| use SADA instead of TeaCache | Flux flow-matching · MS-COCO 2017 | FID: 4.89 -> 1.95 | Helped |
| use SADA instead of AdaptiveDiffusion | SD-2 with Euler · MS-COCO 2017 | Speedup ratio: 1.89× -> 1.81× | Hurt |
| reduce the sampling steps from 50 to 25 or 15 (ablation) | SADA on SD-2 and SDXL with Euler or DPM++ · MS-COCO 2017 | FID: 4.26 to 3.70; 3.76 to 3.26 | Mixed |
| reduce the sampling steps from 50 to 25 or 15 (ablation) | SADA on SD-2 and SDXL with Euler or DPM++ · MS-COCO 2017 | Speedup ratio: 1.81× -> 1.48×; 1.81× -> 1.25× | Hurt |
| use Adams–Moulton updates instead of third-order finite differences (ablation) | SADA step-wise approximation · 50 randomly sampled MS-COCO prompts | Reconstruction error | Helped |
Surprising findings
- With fewer sampling steps (15 vs 50), SADA's fidelity relative to the baseline improves (PSNR rises from 26.25 to 29.34 dB for Euler on SD-2), while speedup drops to about 1.25x.
Limitations
- Evaluation fidelity is measured relative to the unmodified baseline outputs, not against human or reference-quality judgments.
- Speedups are modest (about 1.8-2x) and fall to about 1.25x in the 15-step setting.
- Lagrange interpolation parameters are adjusted by hand for shorter schedules, so some hyperparameter tuning is still needed.
Experimental setup
Datasets: MS-COCO 2017
Benchmarks: MS-COCO 2017
Metrics: PSNR LPIPS FID speedup ratio
Baselines: DeepCache AdaptiveDiffusion TeaCache
Builds on: DeepCache AdaptiveDiffusion DPM-Solver++ ToMe
Closest papers by meaning
- Accelerating Convergence of Score-Based Diffusion Models, Provably (ICML 2024 · similarity 0.90)
- Adaptive Stochastic Coefficients for Accelerating Diffusion Sampling (NeurIPS 2025 · similarity 0.90)
- On Accelerating Diffusion-Based Sampling Processes via Improved Integration Approximation (ICLR 2024 · similarity 0.89)
- Morse: Dual-Sampling for Lossless Acceleration of Diffusion Models (ICML 2025 · similarity 0.89)
- Learning to Solve Generative ODEs Beyond the Linear Span (NeurIPS 2026 · similarity 0.89)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.