Research map / Image generation / Fast diffusion sampling
SenCache: Accelerating Diffusion Model Inference via Sensitivity-Aware Caching
Yasaman Haghighi, Alexandre Alahi
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 |
|---|---|---|---|
| use sensitivity-based criterion instead of MagCache residual-magnitude criterion (aggressive regime) | caching for Wan 2.1 · Wan 2.1 (VBench, fast) | LPIPS: 0.0603 -> 0.0540 (same NFE 21) | Improved |
| use sensitivity-based criterion instead of MagCache (conservative regime) | caching for Wan 2.1 · Wan 2.1 (VBench, slow) | LPIPS: 0.0390 -> 0.0390 | No clear effect |
| use sensitivity criterion instead of TeaCache | caching for Wan 2.1 slow · Wan 2.1 (VBench, slow) | NFE: 33 -> 25 | Improved |
| use sensitivity criterion instead of MagCache | caching for CogVideoX · CogVideoX (VBench) | LPIPS: 0.1952 -> 0.1901 | Improved |
| use sensitivity criterion instead of MagCache | caching for LTX-Video · LTX-Video (VBench) | LPIPS: 0.1795 -> 0.1625 | Improved |
| increase max consecutive cache steps n from 1 to 4 (ablation) | SenCache on Wan 2.1 · Wan 2.1 (ablation n) | NFE: 32 -> 23 | Improved |
| increase n beyond 4 (4 to 7) (ablation) | SenCache on Wan 2.1 · Wan 2.1 (ablation n) | NFE: 23 -> 23 | No clear effect |
| increase n from 4 to 7 (ablation) | SenCache on Wan 2.1 · Wan 2.1 (ablation n) | LPIPS: 0.0558 -> 0.0760 | Hurt |
| increase tolerance epsilon from 0.04 to 0.13 (ablation) | SenCache on Wan 2.1 · Wan 2.1 (ablation eps) | NFE: 25 -> 21 | Improved |
| increase tolerance epsilon from 0.04 to 0.13 (ablation) | SenCache on Wan 2.1 · Wan 2.1 (ablation eps) | LPIPS: 0.0455 -> 0.0513 | Hurt |
| calibration set of 8 videos instead of 4096 (ablation) | SenCache sensitivity estimation · sensitivity profile | profile match | No clear effect |
Closest papers by meaning
- Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model (CVPR 2025 · similarity 0.92)
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.92)
- Cache Me if You Can: Accelerating Diffusion Models through Block Caching (CVPR 2024 · similarity 0.92)
- DeepCache: Accelerating Diffusion Models for Free (CVPR 2024 · similarity 0.92)
- CacheQuant: Comprehensively Accelerated Diffusion Models (CVPR 2025 · similarity 0.92)
- MagCache: Fast Video Generation with Magnitude-Aware Cache (NeurIPS 2025 · similarity 0.91)
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.
