SpecLoR: Spectral Lookahead Rectification for Motion-Coherent Text-to-Video Generation

2026-06-10

AI Summary

Text-to-video generation often suffers from spatiotemporal inconsistencies and motion distortions due to the accumulation of discretization errors during numerical integration. This work proposes a plug-and-play inference method that, for the first time, integrates forward-looking prediction with frequency-domain correction. Specifically, it predicts a clean latent representation and leverages statistical priors derived from the magnitude spectra of natural videos in the frequency domain to perform correction while preserving critical phase information. The corrected latent is then re-noised to continue the latent-space ODE integration. By decoupling noise interference from spatiotemporal dynamics, this approach significantly enhances the physical plausibility and motion coherence of generated videos with only four additional function evaluations (NFE). Substantial reductions in visual artifacts are demonstrated on benchmarks such as Wan2.2.

Abstract

Flow Matching has enabled robust text-to-video generation via latent ODE sampling. However, velocity approximation and numerical discretization errors inevitably accumulate, causing sampling trajectories to drift. Consequently, generated videos often suffer from severe spatiotemporal inconsistencies. Nevertheless, directly correcting these drifted, noisy latents is challenging: (i) timestep-dependent noise obscures reliable structural cues; (ii) spatial interventions risk disrupting intricate local geometry while incurring heavy computational costs. To address this, we propose Spectral Lookahead Rectification (SpecLoR), a plug-and-play inference method that bypasses noise via lookahead prediction, and circumvents spatiotemporal entanglement by shifting corrections to the frequency domain, where universal statistical priors of natural videos are readily available. First, during early sampling stages, SpecLoR looks ahead to estimate the clean latent $z_{t,0}$ and computes its 3D spatiotemporal spectrum. Next, SpecLoR rectifies the amplitude spectrum to match the prior, leaving the phase intact. Finally, the corrected state is re-noised to resume ODE integration. Experiments on Wan2.2 demonstrate that SpecLoR significantly reduces physical artifacts and enhances motion coherence across multiple benchmarks with minimal computational overhead (4 additional NFEs).

Problem

• text-to-video generation
• spatiotemporal inconsistency
• sampling drift
• flow matching
• motion coherence

Innovation

• Spectral Lookahead Rectification
• Flow Matching
• Text-to-Video Generation
• Frequency Domain Correction
• Motion Coherence