Research map / Image generation / Fast diffusion sampling
SpecDiff: Accelerating Diffusion Model Inference with Self-Speculation
Jiayi Pan, Jiaming Xu, Yongkang Zhou, Guohao Dai
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 |
|---|---|---|---|
| add token hierarchy algorithm (ablation) | SpecDiff · FID | FID: reduces the FID by 18% | Improved |
| add token hierarchy algorithm (ablation) | SpecDiff · Clip Score | Clip Score: increases the clip score by 2.8% | Improved |
| add feature prediction mechanism (ablation) | SpecDiff · FID | FID: further reduces the FID by 3.6% | Improved |
| add feature prediction mechanism (ablation) | SpecDiff · Clip Score | Clip Score: further increases the clip score by 0.35% | Improved |
| add feature prediction mechanism (ablation) | SpecDiff · FLUX speedup | speedup: 2.36x | Improved |
| add token hierarchy algorithm (ablation) | SpecDiff · FLUX speedup | speedup: further accelerates by 1.34x; eventually 3.17x | Improved |
| increase speculation steps from 2 to 4 (99% cached) (ablation) | SpecDiff · SD3 | FID: 29.75 -> 29.64 | No clear effect |
| increase speculation steps from 2 to 4 (99% cached) (ablation) | SpecDiff · SD3 | speedup: 2.80x -> 2.25x | Hurt |
| increase speculation steps from 2 to 4 (92% cached) (ablation) | SpecDiff · SD3 | FID: 29.52 -> 29.12 | No clear effect |
| increase speculation steps from 2 to 4 (55% cached) (ablation) | SpecDiff · SD3 | speedup: 1.61x -> 1.33x | Hurt |
| replace RAS 87.5% with SpecDiff 99% | RAS · SD3 | FID: 40.92 -> 29.75 | Improved |
| replace RAS 87.5% with SpecDiff 99% | RAS · SD3.5 | FID: 36.87 -> 30.08 | Improved |
Closest papers by meaning
- Beyond Fixed Formulas: Data-Driven Linear Predictor for Efficient Diffusion Models (CVPR 2026 · similarity 0.91)
- Diffusion on Demand: Selective Caching and Modulation for Efficient Generation (NeurIPS 2025 · similarity 0.90)
- Accelerating Diffusion Models with Parallel Sampling: Inference at Sub-Linear Time Complexity (NeurIPS 2024 · similarity 0.87)
- Fast and Memory-Efficient Video Diffusion Using Streamlined Inference (NeurIPS 2024 · similarity 0.87)
- Accelerating Diffusion Transformers with Token-wise Feature Caching (ICLR 2025 · similarity 0.87)
- Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching (NeurIPS 2024 · similarity 0.87)
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.
