Research map / Video / Video generation
NaviCache: Test-Time Self-Calibration Caching for Video Generation
Zheqi Lv, Zhibo Zhu, Jinke Wang, Qi Tian, Shengyu Zhang, Zhengyu Chen, Chengxi Zang, Zhou Zhao, Fei Wu
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 NaviCache-mid instead of EasyCache | HunyuanVideo caching · HunyuanVideo VBench prompts | PSNR: 32.53 -> 32.65 (speedup 2.15x -> 2.17x) | Improved |
| use NaviCache-slow instead of TeaCache/MagCache | Wan 2.1-1.3B · Wan 2.1-1.3B VBench prompts | PSNR: 22.79/23.33 -> 25.10 | Improved |
| use NaviCache instead of TeaCache on VBench score | Wan 2.1-1.3B · Wan 2.1-1.3B VBench | VBench score: 80.67 vs 80.21 (fast) | No clear effect |
| increase Initial Alignment steps from 5 to 15 (ablation) | NaviCache · Wan 2.1-1.3B | PSNR: 22.73 -> 24.31 | Improved |
| increase Initial Alignment steps from 5 to 15 (ablation) | NaviCache · Wan 2.1-1.3B | latency (s): 92.31 -> 112.49 | Hurt |
| decrease process noise Q from 0.05I to 0.005I (ablation) | NaviCache (tau 0.04I) · Wan 2.1-1.3B | PSNR: 25.10 -> 30.33 | Improved |
| decrease process noise Q from 0.05I to 0.005I (ablation) | NaviCache (tau 0.04I) · Wan 2.1-1.3B | latency (s): 115.86 -> 153.29 | Hurt |
| decrease measurement noise R from 0.05I to 0.005I (ablation) | NaviCache (tau 0.07I) · Wan 2.1-1.3B | PSNR: 23.46 -> 22.36 | Hurt |
| increase threshold tau from 0.04I to 0.07I (ablation) | NaviCache · Wan 2.1-1.3B | speedup / LPIPS: LPIPS 0.0686 -> 0.0956 | Mixed |
| test-time self-calibration with Kalman-style filter vs offline calibration cost | NaviCache · Auxiliary overhead | calibration time (s): 15191 (TeaCache) -> 0 | 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.
- MagCache — MagCache: Fast Video Generation with Magnitude-Aware Cache (NeurIPS 2025) · better on 4 of 6 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
- LearniBridge: Learnable Calibration of Feature Caching for Diffusion Models Acceleration (ICML 2026 · similarity 0.90)
- SenCache: Accelerating Diffusion Model Inference via Sensitivity-Aware Caching (CVPR 2026 · similarity 0.89)
- MagCache: Fast Video Generation with Magnitude-Aware Cache (NeurIPS 2025 · similarity 0.89)
- Accelerating Rectified Flow Models via Trajectory-Aware Caching (NeurIPS 2026 · similarity 0.88)
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.87)
- Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model (CVPR 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.
