In-Distribution Forcing for Long Video Generation at Test Time

2026-10-02

AI Summary

This study addresses the issues of color-texture drift and motion degradation in autoregressive video diffusion models when generating long videos, which stem from out-of-distribution KV caches. To tackle this, we propose ID-Forcing, a framework that introduces a novel self-caching mechanism to align cache states with training configurations during inference, thereby preventing KV entries from drifting out of distribution at their source. Notably, this approach extends generation duration to the minute scale without requiring retraining. Experiments demonstrate that ID-Forcing effectively suppresses drift artifacts while maintaining competitive performance on standard benchmarks. Both quantitative evaluations and user studies confirm its superiority over existing methods.

Abstract

Modern autoregressive (AR) video diffusion models excel at short-horizon video generation, yet generating long videos remains challenging due to drifting, where colors and textures shift, and motion dynamics decay. Existing works primarily rely on KV conditioning, which selects or modifies cached key-value (KV) entries to mitigate drifting. However, we observe that KV conditioning alone is insufficient as it assumes cached KV entries remain in-distribution. This assumption fails beyond the training horizon: nothing constrains the construction of KV entries during rollout, giving rise to the KV-provenance problem where cached entries themselves become out-of-distribution (OOD). To address this, we propose In-Distribution Forcing (ID-Forcing), a test-time framework that aligns both KV caching and KV conditioning with training configurations. Its key mechanism, self-caching, prevents OOD KV entries at their source. Each chunk is cached without attending to prior KV entry, keeping the rolling window exactly in-distribution. Consequently, ID-Forcing seamlessly extends short-horizon models to minute-scale video generation. Extensive evaluations show that our method remains competitive on standard video generation benchmark while substantially outperforming prior work in mitigating drifting, as validated by both our drift metrics and a user study.

Problem

• long video generation
• drifting
• KV-provenance problem
• out-of-distribution
• autoregressive video diffusion models

Innovation

• In-Distribution Forcing
• Long Video Generation
• Self-Caching
• Autoregressive Diffusion Models
• KV-Provenance Problem