Technical Summary: NaviCache

Problem Statement

Video Diffusion Models (VDMs), such as HunyuanVideo, Wan, and Open-Sora, have achieved significant advancements in high-fidelity content generation. However, their iterative sampling process imposes a massive computational tax, requiring dozens of forward passes through large architectures, which hinders real-time deployment.

Existing acceleration strategies fall into two categories, both with significant limitations:

1. Offline Calibration-Dependent Methods: Approaches like TeaCache and MagCache rely on statistical priors or polynomial fitting derived from curated calibration datasets. These suffer from high calibration costs, dependency on specific data distributions, and performance degradation when faced with distribution shifts between calibration and inference.
2. Offline Calibration-Free Methods: Approaches like EasyCache eliminate the need for calibration datasets but rely on instantaneous zero-order approximations (Zero-Order Hold). These methods assume the feature change ratio remains constant between steps, failing to capture the intrinsic momentum of the diffusion trajectory. This leads to susceptibility to observational noise and significant errors in high-turbulence regions, resulting in either visual artifacts (over-skipping) or insufficient acceleration (under-skipping).

The core challenge is to develop a method that avoids offline calibration overhead while accurately modeling the non-stationary, momentum-driven dynamics of feature evolution during the diffusion process.

Methodology: NaviCache

The authors propose NaviCache, a plug-and-play, test-time self-calibration framework that re-conceptualizes feature evolution as an Inertial Navigation System (INS) problem. Instead of estimating absolute changes, NaviCache models the relative coupling between input and output variations as a state-space tracking problem.

Core Components

1. State-Space Modeling:
   The method formulates the feature change ratio ($\phi_t = \Delta O_t / \Delta I_t$) as a hidden state $r_t$ within a Linear Gaussian State-Space Model (LGSSM). It explicitly models the uncertainty and kinematic inertia of this ratio, rather than treating it as a static constant.

2. Initial Alignment Phase:
   Recognizing that diffusion processes exhibit chaotic, non-stationary dynamics in early timesteps (high noise), NaviCache employs a strategic "Initial Alignment" phase. For the first $N_{align}$ steps, full computation is performed without skipping. This allows the system to compute an unbiased initialization for the state mean ($\hat{r}_0$) and error covariance ($P_0$), ensuring the estimator engages "navigation mode" with a calibrated understanding of the feature manifold.

3. Dual-State Estimation Engine:
   During the main diffusion phase, NaviCache utilizes a recursive Bayesian filter to track two sufficient statistics: the state estimate ($\hat{r}_t$) and the estimation uncertainty ($P_t$). The process involves two stages:

   • Prior Projection: The system projects the state and uncertainty forward, incorporating process noise ($Q_t$) to model the intrinsic drift or "momentum" of the feature ratio.
   • Observational Rectification: When the accumulated error exceeds a safety threshold, a full computation is triggered to obtain a ground-truth measurement ($z_t$). This measurement is fused with the prior prediction using a Calibrated Fusion Factor (CFF), $K_t$, derived to minimize posterior uncertainty. This mechanism acts as an intelligent gatekeeper, prioritizing historical inertia in stable regimes and aligning instantly with new observations in turbulent regimes.

4. Uncertainty-Aware Skipping Mechanism:
   The decision to skip or update is governed by an accumulated error metric ($E_{acc}$). If $E_{acc}$ remains below a fidelity threshold ($\tau$), the system skips the computation (Dead Reckoning). If the threshold is breached, the system performs a full update (Measurement Fix), re-anchoring the trajectory and resetting the error accumulator.

Key Contributions

The paper outlines four primary contributions:

• Theoretical Reformulation: The authors establish the first formal mapping between diffusion dynamics and Inertial Navigation System (INS) theory, reformulating VDM feature evolution as a state-space model.
• NaviCache Framework: They propose an offline calibration-free framework featuring a dual-state estimation engine and an initial alignment phase to accurately track feature manifolds without pre-training or calibration datasets.
• Theoretical Guarantees: The paper provides a theoretical proof that NaviCache achieves a lower estimation error bound than the zero-order heuristics used in existing calibration-free methods, provided process and measurement noise exist.
• Empirical Performance: Extensive experiments demonstrate that NaviCache consistently outperforms baselines across multiple models (Wan, HunyuanVideo, Open-Sora) in both inference efficiency and visual fidelity.

Experimental Results

Experiments were conducted on three representative VDMs: Wan 2.1, HunyuanVideo, and Open-Sora 1.2, evaluated on the VBench dataset.

• Efficiency vs. Quality Trade-off: NaviCache offers three configurations (fast, mid, slow). In comparable latency envelopes, NaviCache consistently demonstrates superior visual retention over offline calibration-free methods like EasyCache. For instance, on HunyuanVideo, the "mid" mode achieved a 2.17× speedup with a PSNR of 32.65, outperforming EasyCache's 2.15× speedup (PSNR 32.53).
• Comparison with Calibration-Dependent Methods: NaviCache exhibits remarkable fidelity compared to offline calibration methods. On Wan 2.1, the "slow" mode achieved a PSNR of 25.10, significantly higher than TeaCache (22.79) and MagCache (23.33). While TeaCache occasionally maintains a marginal lead in VBench scores (likely due to prompt-specific fitting), NaviCache provides superior pixel-level and perceptual reconstruction (lower LPIPS, higher SSIM).
• Visual Fidelity: Case studies reveal that existing methods often suffer from temporal inconsistencies, ghosting, and structural distortions (e.g., deformed hands or objects). NaviCache effectively mitigates these issues, maintaining structural integrity and high-frequency details even during rapid movement.
• Ablation Studies:
  • Initial Alignment: Increasing alignment steps improves fidelity but increases latency; a 20% ratio was found to be an optimal balance.
  • Noise Parameters: Lower process noise ($Q$) improves fidelity at the cost of latency, while lower measurement noise ($R$) reduces latency but risks visual retention.
  • Scalability: The method maintains stable acceleration (1.81×–1.91×) across varying resolutions (480p/720p) and frame counts (51/102).
• Overhead: NaviCache requires zero offline calibration time. The test-time computational overhead for the skipping justification is negligible (approx. 0.0109s).

Significance and Claims

The paper claims that NaviCache bridges the fundamental domain gap between control theory and video generation. By modeling feature evolution as a structured trajectory with underlying momentum rather than random jitters, NaviCache enables plug-and-play test-time self-calibration.

The authors assert that this approach eliminates the prohibitive calibration costs and distribution shift vulnerabilities of offline methods while overcoming the accuracy limitations of zero-order approximations. Theoretically grounded in error-bounded computation skipping, NaviCache is presented as a principled solution for the efficient deployment of large-scale generative models, contributing to the goal of "Green AI" by significantly lowering computational and energy barriers for real-time video synthesis.