Pathwise Test-Time Correction for Autoregressive Long Video Generation
This addresses a specific bottleneck in long video synthesis for applications requiring real-time generation, representing an incremental improvement over existing test-time optimization methods.
The paper tackles error accumulation in long video generation with autoregressive diffusion models by introducing Test-Time Correction (TTC), which uses the initial frame as a reference to calibrate intermediate states during sampling. The method extends generation lengths with negligible overhead and matches the quality of training-based methods on 30-second benchmarks.
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.