Saved in:
Bibliographic Details
Main Authors: Xiang, Xunzhi, Duan, Zixuan, Zhang, Guiyu, Zhang, Haiyu, Gao, Zhe, Wu, Junta, Zhang, Shaofeng, Wang, Tengfei, Fan, Qi, Guo, Chunchao
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2602.05871
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914380508037120
author Xiang, Xunzhi
Duan, Zixuan
Zhang, Guiyu
Zhang, Haiyu
Gao, Zhe
Wu, Junta
Zhang, Shaofeng
Wang, Tengfei
Fan, Qi
Guo, Chunchao
author_facet Xiang, Xunzhi
Duan, Zixuan
Zhang, Guiyu
Zhang, Haiyu
Gao, Zhe
Wu, Junta
Zhang, Shaofeng
Wang, Tengfei
Fan, Qi
Guo, Chunchao
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2602_05871
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Pathwise Test-Time Correction for Autoregressive Long Video Generation
Xiang, Xunzhi
Duan, Zixuan
Zhang, Guiyu
Zhang, Haiyu
Gao, Zhe
Wu, Junta
Zhang, Shaofeng
Wang, Tengfei
Fan, Qi
Guo, Chunchao
Computer Vision and Pattern Recognition
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.
title Pathwise Test-Time Correction for Autoregressive Long Video Generation
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2602.05871