Pathwise Test-Time Correction for Autoregressive Long Video Generation

TL;DR

This paper introduces Test-Time Correction (TTC), a training-free method that uses the initial frame as a reference to improve long-sequence autoregressive video generation, reducing drift and maintaining quality.

Contribution

The paper proposes TTC, a novel training-free correction technique that stabilizes long video generation by calibrating intermediate states using the initial frame as an anchor.

Findings

TTC extends video generation length with minimal overhead.

TTC matches the quality of training-based methods on 30-second benchmarks.

TTC effectively mitigates drift in extended sequences.

Abstract

Distilled autoregressive diffusion models facilitate real-time short video synthesis but suffer from severe error accumulation during long-sequence generation. While existing Test-Time Optimization (TTO) methods prove effective for images or short clips, we identify that they fail to mitigate drift in extended sequences due to unstable reward landscapes and the hypersensitivity of distilled parameters. To overcome these limitations, we introduce Test-Time Correction (TTC), a training-free alternative. Specifically, TTC utilizes the initial frame as a stable reference anchor to calibrate intermediate stochastic states along the sampling trajectory. Extensive experiments demonstrate that our method seamlessly integrates with various distilled models, extending generation lengths with negligible overhead while matching the quality of resource-intensive training-based methods on 30-second benchmarks.