Pathwise Test-Time Correction for Autoregressive Long Video Generation

2026-02-05

AI Summary

This work addresses the drift problem in long video generation with distilled autoregressive diffusion models, which arises from error accumulation during sequential sampling. To mitigate this issue, the authors propose a training-free, test-time path-level correction method that leverages the initial frame as a stable anchor to dynamically recalibrate intermediate stochastic states along the sampling trajectory. This approach effectively suppresses generation drift without requiring reward signals or parameter fine-tuning. Despite its minimal computational overhead, the method achieves generation quality on par with costly training-based alternatives on a 30-second video benchmark, significantly enhancing both the achievable generation length and temporal stability.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Natural Language Processing: Generation
• Machine Learning: Deep Generative Models & Autoencoders

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

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.

Problem

• autoregressive video generation
• error accumulation
• test-time optimization
• long-sequence generation
• distilled diffusion models

Innovation

• Test-Time Correction
• autoregressive video generation
• distilled diffusion models
• error accumulation mitigation
• training-free adaptation