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

2026-06-25

AI Summary

Video diffusion models suffer from high inference costs, and existing acceleration methods either rely on offline calibration or are sensitive to noise due to neglecting the momentum characteristics of diffusion trajectories. This work proposes NaviCache, the first framework to introduce inertial navigation principles into video generation acceleration by modeling feature evolution as a navigation problem. NaviCache employs a dual-state estimation architecture, an initial alignment mechanism, time-varying noise scheduling, and uncertainty-aware measurement updates to construct a plug-in, test-time self-calibrating caching framework that enables calibration-free, error-bounded computation skipping. Extensive experiments on HunyuanVideo, Wan, and Open-Sora demonstrate that NaviCache significantly improves the accuracy of skip decisions and achieves superior overall acceleration performance.

Abstract

Video Diffusion Models (VDMs) is constrained by immense computational costs. While offline calibration-based acceleration suffers from calibration data dependency, prohibitive calibration duration, and susceptibility to distribution shifts, offline calibration-free methods eliminate these hurdles. However, since they rely on instantaneous zero-order approximations where the mapping between input and output differences varies in real-time, they are susceptible to observational noise and ignore the intrinsic momentum within the diffusion trajectory. In this paper, we propose NaviCache, a plug-and-play test-time self-calibration method re-conceptualizing feature evolution as an Inertial Navigation System (INS) problem. NaviCache bridges the fundamental domain gap and the non-stationary nature of diffusion by modeling the relative coupling between input and output variations. We introduce a dual-state estimation architecture that adaptively tracks the feature change ratio and its latent drift, initialized via a specialized Initial Alignment phase. By integrating a time-dependent noise schedule with an uncertainty-aware Measurement Update mechanism, NaviCache provides a theoretically grounded mechanism for error-bounded computation skipping. Extensive experiments on the HunyuanVideo, Wan, and Open-Sora series demonstrate that NaviCache exhibits more accurate error judgment for computation skipping and achieves outstanding comprehensive performance.

Problem

• Video Diffusion Models
• Computational Cost
• Test-Time Acceleration
• Distribution Shift
• Diffusion Trajectory

Innovation

• test-time self-calibration
• video diffusion models
• inertial navigation system
• computation skipping
• uncertainty-aware update