Research map / Video / Video generation
PreciseCache: Precise Feature Caching for Efficient and High-fidelity Video Generation
Jiangshan Wang, Kang Zhao, Jiayi Guo, Jiayu Wang, Hang Guo, Chenyang Zhu, Xiu Li, Xiangyu Yue
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 TeaCache with PreciseCache-Turbo (LFD-based step caching) | feature caching for video diffusion · Wan2.1-14B 720P | speedup and LPIPS: TeaCache 1.94x / LPIPS 0.1012 -> Turbo 2.15x / 0.0633 | Improved |
| PreciseCache-Turbo vs TeaCache on Open-Sora 1.2 | feature caching · Open-Sora 1.2 | VBench: TeaCache 78.23% -> Turbo 78.49% (base 78.79%) | No clear effect |
| add BlockCache on top of Turbo (Flash) (ablation) | PreciseCache · Wan2.1-14B | speedup: 2.15x -> 2.63x with VBench 83.52% -> 83.43% | Improved |
| use 1x8x8 downsample for LFD estimate instead of 2x4x4 (ablation) | LFCache · Wan2.1-14B | VBench / LPIPS: VBench 83.52% -> 83.18%; LPIPS 0.0793 -> 0.1946 | Hurt |
| use 4x4x4 downsample instead of 2x4x4 (ablation) | LFCache · Wan2.1-14B | VBench / LPIPS: VBench 83.52% -> 83.02%; LPIPS 0.0793 -> 0.1875 | Hurt |
| use small 1x2x2 downsample instead of 2x4x4 (ablation) | LFCache · Wan2.1-14B | latency: speedup 2.18x -> 0.98x | Hurt |
| cache residuals instead of final predictions (ablation) | LFCache · Wan2.1-14B | VBench / LPIPS: VBench 83.52% -> 83.50%; LPIPS 0.0793 -> 0.0791 | No clear effect |
| use TaylorSeer reuse instead of reusing prediction (ablation) | LFCache · Wan2.1-14B | VBench / LPIPS: VBench 83.52% -> 83.54%; LPIPS 0.0793 -> 0.0801 | No clear effect |
| scale PreciseCache-Turbo from 1 to 8 GPUs (DSP) | PreciseCache · Wan2.1 1080P | latency (s): 1 GPU: 3326 -> 1330; 8 GPUs: 459 -> 229 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FasterCache — FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025) · better on 9 of 9 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.89)
- Flow Caching for Autoregressive Video Generation (ICLR 2026 · similarity 0.89)
- Motion-Aware Caching for Efficient Autoregressive Video Generation (ICML 2026 · similarity 0.89)
- MagCache: Fast Video Generation with Magnitude-Aware Cache (NeurIPS 2025 · similarity 0.89)
- LeMiCa: Lexicographic Minimax Path Caching for Efficient Diffusion-Based Video Generation (NeurIPS 2025 · similarity 0.87)
- Accelerating Diffusion Transformers with Token-wise Feature Caching (ICLR 2025 · similarity 0.86)
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.
