Research map / Image generation / Fast diffusion sampling
Adaptive Stochastic Coefficients for Accelerating Diffusion Sampling
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.
Proposes AdaSDE, a single-step SDE solver for few-step diffusion sampling that learns one stochastic coefficient per step via lightweight trajectory distillation, reporting SOTA FIDs at 3–9 NFE.
Problem
Few-step (low-NFE) diffusion sampling trades off speed and quality; ODE solvers accumulate gradient error while SDE solvers suffer discretization error when step budgets are small.
Approach
Introduce a learnable per-step stochastic coefficient gamma_i that controls forward noise injection followed by a deterministic backward ODE step, with the coefficients (plus time-shift and scale parameters) optimized stagewise by aligning student states to refined teacher trajectories.
Key results
- At 5 NFE, AdaSDE reaches FID 4.18 on CIFAR-10 versus 7.59 for AMED-Solver.
- At 5 NFE, AdaSDE reaches FID 8.05 on FFHQ 64x64 versus 14.80 for AMED-Solver.
- On Stable Diffusion v1.5 (MS-COCO, 512x512) at 10 NFE, AdaSDE reaches FID 12.68 versus 13.24 for AMED-Plugin.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AdaSDE instead of AMED-Solver | diffusion sampling · CIFAR-10 32×32, NFE=5 | FID: 7.59 -> 4.18 | Helped |
| use AdaSDE instead of AMED-Solver | diffusion sampling · FFHQ 64×64, NFE=5 | FID: 14.80 -> 8.05 | Helped |
| use AdaSDE instead of AMED-Solver | diffusion sampling · ImageNet 64×64, NFE=5 | FID: 10.74 -> 6.90 | Helped |
| use AdaSDE instead of AMED-Solver | diffusion sampling · LSUN Bedroom 256×256, NFE=3 | FID: 58.21 -> 18.03 | Helped |
| use AdaSDE instead of AMED-Plugin | Stable Diffusion v1.5 · MS-COCO 512×512, NFE=8 | FID: 13.96 -> 13.39 | Helped |
| remove learned stochastic coefficient γₙ (ablation) | AdaSDE · CIFAR-10 32×32, NFE=3 | FID: 12.62 -> 13.32 | Hurt |
| remove learned stochastic coefficient γₙ (ablation) | AdaSDE · CIFAR-10 32×32, NFE=5 | FID: 4.18 -> 4.36 | Hurt |
| remove learned stochastic coefficient γₙ (ablation) | AdaSDE · FFHQ 64×64, NFE=3 | FID: 23.80 -> 25.85 | Hurt |
| remove learned stochastic coefficient γₙ (ablation) | AdaSDE · FFHQ 64×64, NFE=5 | FID: 8.04 -> 8.11 | Hurt |
| remove learned stochastic coefficient γₙ (ablation) | AdaSDE · MS-COCO 512×512, NFE=4 | FID: 30.89 -> 37.23 | Hurt |
| remove learned stochastic coefficient γₙ (ablation) | AdaSDE · MS-COCO 512×512, NFE=6 | FID: 13.79 -> 16.34 | Hurt |
| remove learned stochastic coefficient γₙ (ablation) | AdaSDE · MS-COCO 512×512, NFE=10 | FID: 12.68 -> 12.82 | Hurt |
Surprising findings
- Removing the learned stochastic coefficient gamma only modestly degrades FID on CIFAR-10 and FFHQ at higher NFE, while the gap is large on SD v1.5 at low NFE.
- At NFE=3 on CIFAR-10, the polynomial time schedule slightly beats the uniform schedule (11.61 vs 12.62).
Limitations
- Large step sizes with strong stochastic injection (high gamma) can cause error amplification and instability, so gamma must be clipped or the schedule made conservative.
- Per-step distribution resets and geometric alignment break linear-recurrence assumptions of multistep and predictor-corrector methods (iPNDM, UniPC).
Experimental setup
Datasets: CIFAR-10 FFHQ 64x64 ImageNet 64x64 LSUN Bedroom MS-COCO
Benchmarks: CIFAR-10 FFHQ 64x64 ImageNet 64x64 LSUN Bedroom 256x256 MS-COCO 512x512
Metrics: FID 1-Wasserstein distance
Baselines: DPM-Solver++(3M) UniPC iPNDM DDIM Heun DPM-Solver-2 DPM-Plugin AMED-Solver DPM-Solver++(2M) DPM-Solver-v3 AMED-Plugin
Builds on: AMED-Solver DPM-Solver EDM Restart sampling iPNDM
Closest papers by meaning
- On Accelerating Diffusion-Based Sampling Processes via Improved Integration Approximation (ICLR 2024 · similarity 0.92)
- Distilling ODE Solvers of Diffusion Models into Smaller Steps (CVPR 2024 · similarity 0.91)
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.91)
- Accelerating Convergence of Score-Based Diffusion Models, Provably (ICML 2024 · similarity 0.91)
- SADA: Stability-guided Adaptive Diffusion Acceleration (ICML 2025 · similarity 0.90)
- On the Trajectory Regularity of ODE-based Diffusion Sampling (ICML 2024 · similarity 0.90)
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.