AI

In-Distribution Forcing for Long Video Generation at Test Time

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 09:18:24 UTC.

Briefings

This research presents In-Distribution Forcing (ID-Forcing), a test-time framework for autoregressive video diffusion models that extends short-horizon generation to minute-scale videos by aligning KV caching and conditioning with training configurations. The method uses self-caching to prevent out-of-distribution key-value entries during rollout, effectively mitigating drifting and motion decay. Evaluated on standard benchmarks, the approach remains competitive while significantly outperforming prior work in reducing visual drift.

Key points

• The proposed method, In-Distribution Forcing, uses self-caching to align KV caching and conditioning with training configurations, preventing out-of-distribution key-value entries during rollout.
• Extensive evaluations show the method remains competitive on standard video generation benchmarks while substantially outperforming prior work in mitigating drifting and motion decay.

Context

Modern autoregressive video diffusion models excel at short-horizon generation but struggle with long videos due to drifting and motion decay. Existing KV conditioning methods assume cached entries stay in-distribution, which fails beyond the training horizon. This study addresses the KV-provenance problem by proposing ID-Forcing to extend short-horizon capabilities to minute-scale generation.

Limitations

The study relies on internal drift metrics and user study validation rather than independent external benchmarks. The evidence is limited to the authors' internal evaluations and does not include independent third-party validation or long-term stability testing beyond the reported metrics.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 19/10/2026, 15:15:02 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiad7axebzfljy4lbjpfy6qp423mcozealfrjwrtmcayo2k7qtvu7q