FuguReport

NaviCache: Test-Time Self-Calibration Caching for Video Generation

Authors Zheqi Lv, Zhibo Zhu, Jinke Wang, Qi Tian, Shengyu Zhang, Zhengyu Chen, Chengxi Zang, Zhou Zhao, Fei Wu
Affiliations Cornell University / Zhejiang University / Tencent / MeiTuan LongCat
Categories Method / Model Calibration / Self-calibration for video diffusion models, Application / Video Generation / Enhancing video diffusion model performance, Theory / Error Analysis / Theoretical error bounds and skip mechanisms
License CC BY 4.0

Abstract Overview

NaviCache is a plug-and-play, test-time self-calibration caching method for accelerating video diffusion models. The paper reframes the relationship between input feature changes and output responses as a state-space tracking problem inspired by inertial navigation systems, using an initial alignment stage, dual-state estimation, and uncertainty-aware update decisions to control computation skipping. This design is intended to avoid the calibration-data dependence and distribution-shift issues of offline calibration-based methods while improving on zero-order calibration-free heuristics. The paper also provides a theoretical analysis claiming a lower steady-state estimation error bound than zero-order heuristics, and evaluates the method on Wan 2.1, HunyuanVideo, and Open-Sora 1.2.

Novelty

The paper’s main novelty is treating video diffusion feature evolution as an inertial-navigation-style state estimation problem rather than using offline-fitted mappings or instantaneous zero-order approximations. It introduces a test-time self-calibration mechanism with initial alignment, recursive dual-state estimation, and uncertainty-triggered measurement updates for skip decisions.

Results

Across Wan 2.1, HunyuanVideo, and Open-Sora 1.2, NaviCache demonstrates a strong efficiency-quality trade-off relative to prior acceleration baselines. For example, on HunyuanVideo, NaviCache-mid achieves a 2.17× speedup with PSNR 32.65, SSIM 0.9256, and LPIPS 0.0571, slightly improving on EasyCache at a similar latency envelope; on Wan 2.1, NaviCache-slow reaches PSNR 25.10, SSIM 0.8638, and LPIPS 0.0686 at 1.86× speedup. The method requires no offline calibration and reports a negligible skip-justification overhead of 0.0109 seconds.

Key Points

  1. NaviCache replaces zero-order skip prediction with online state-space estimation of the output-to-input change ratio, incorporating an initial alignment phase and recursive uncertainty tracking.
  2. The paper provides a theoretical proof that NaviCache achieves a lower steady-state estimation error variance than zero-order heuristics, such as EasyCache, under positive process and measurement noise assumptions.
  3. Experiments across three video diffusion model families confirm that NaviCache preserves visual quality better than prior calibration-free baselines at similar acceleration levels while avoiding offline calibration costs.

References

This page was created using generative AI such as GPT-5, Claude Opus 4, Gemini 3, Gemini 3.1 Flash Image, and their higher-end successor versions. No guarantee can be made regarding its contents.