Research map / Image generation / Fast diffusion sampling
Adaptive Spectral Feature Forecasting for Diffusion Sampling Acceleration
Jiaqi Han, Juntong Shi, Puheng Li, Haotian Ye, Qiushan Guo, Stefano Ermon
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Taylor forecaster with Chebyshev spectral forecaster | TaylorSeer at high speedup · DrawBench FLUX.1 (~4x) | PSNR: 20.24 -> 22.21 PSNR with 4.14x -> 4.79x speedup | Improved |
| replace Taylor forecaster with Chebyshev spectral forecaster | TaylorSeer on Wan2.1-14B · VBench Wan2.1-14B | PSNR: 17.24 -> 21.24 | Improved |
| use adaptive scheduling instead of fixed interval (ablation) | Spectrum on SD3.5-Large · DrawBench SD3.5-Large | PSNR: 11.34 -> 15.68 | Improved |
| use adaptive scheduling instead of fixed interval (ablation) | Taylor forecaster on SD3.5-Large · DrawBench SD3.5-Large | PSNR: 10.40 -> 13.25 | Improved |
| use adaptive scheduling instead of fixed interval (ablation) | Spectrum on Wan2.1-14B · VBench Wan2.1-14B | VBench quality score: PSNR 20.65 -> 21.24; VBench 82.32 -> 82.21 | Mixed |
| cache only the final attention block instead of every module (ablation) | Spectrum on FLUX.1 · FLUX.1 N=8 | PSNR / latency: PSNR 19.34 -> 19.66; latency 15.44 -> 5.59 s | Improved |
| cache only the final attention block in TaylorSeer (ablation) | TaylorSeer on FLUX.1 · FLUX.1 N=8 | PSNR: 17.93 -> 18.61 | Improved |
| set ridge regularization to 1e-3 or 10 instead of 0.1 (ablation) | Spectrum on FLUX.1 · FLUX.1 lambda sweep | image quality | Hurt |
| increase Chebyshev degree M from 2 to 4 (ablation) | Spectrum on FLUX.1 · FLUX.1 degree sweep | sample quality | Improved |
| increase Chebyshev degree M from 4 to 6 (ablation) | Spectrum on FLUX.1 · FLUX.1 degree sweep | sample quality | No clear effect |
| increase Taylor order from 1 to 2 (ablation) | TaylorSeer at high speedup · DrawBench FLUX.1 (N=6) | PSNR: 20.24 -> 17.41 | Hurt |
Closest papers by meaning
- Beyond Fixed Formulas: Data-Driven Linear Predictor for Efficient Diffusion Models (CVPR 2026 · similarity 0.92)
- Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · similarity 0.90)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.90)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.90)
- PFDiff: Training-Free Acceleration of Diffusion Models Combining Past and Future Scores (ICLR 2025 · 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.
