In-Distribution Forcing for Long Video Generation at Test Time
Jeongwoo Shin, Youngyoon Choi, Sangwoo Jo, Hyunmog Kim, Sungjoon Choi, Joonseok Lee, Jaewoong Choi, Jaemoo Choi
Seoul National University · Korea University · Sungkyunkwan University · Georgia Institute of Technology
Published October 2, 2026 · on the feed October 6, 2026 · ▲ 49 · arXiv 2610.03120
The 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.
The Brief
2:17A short podcast summary of the paper. Written and voiced by AI.
Transcript
In-Distribution Forcing for Long Video Generation at Test Time. By Jeongwoo Shin and seven others, from Seoul National University, Korea University, and two other institutions. Published on the second of October. Its code repository is worked out: about forty source files, with a readme.
Autoregressive video diffusion models are good at short clips. Push them toward long videos and they drift: colors and textures shift, and motion dynamics decay. Most existing remedies work through KV conditioning, which selects or modifies the cached key-value entries the model attends to, so that the cache steers generation back on course. The authors observe that this is not enough on its own, because it quietly assumes the cached entries are themselves in-distribution.
That assumption breaks beyond the training horizon. During rollout, nothing constrains how KV entries are constructed in the first place, and so the cached entries can become out-of-distribution. The authors call this the KV-provenance problem. Conditioning on a cache is only as good as the cache, and if its entries were built under conditions the model never saw in training, choosing among them more carefully does not repair them.
In-Distribution Forcing, or ID-Forcing, is a test-time framework built around that diagnosis. Rather than adjusting only how the cache is used, it aligns both halves with the configurations the model was trained under: KV caching as well as KV conditioning. The target is where the bad entries come from, not what happens after they arrive.
Its key mechanism is called self-caching. Each chunk is cached without attending to any prior KV entry, and that is what keeps the rolling window exactly in-distribution. Out-of-distribution entries are prevented at their source rather than managed downstream. The consequence the authors draw is that short-horizon models extend, seamlessly, to minute-scale video generation.
Across extensive evaluations, the method stays competitive on a standard video generation benchmark while substantially outperforming prior work at mitigating drift. That second claim is supported in two ways: by the authors' own drift metrics, and by a user study.
The fix sits one step upstream of where earlier work had been looking. Prior methods asked which cached entries to trust. ID-Forcing asks how those entries came to exist, and arranges for each one to be built the way training built it.
Paper Quest · 7 chapters · about 36 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A beautiful five-second clip, and then the slow slide. Where does it start?
You generate video chunk by chunk, each new chunk attending to stored keys and values from earlier ones. Short clips look great. Past a minute, colors and textures shift and motion fades.
The road ahead
- 1 · The problem ~7 min Long generated videos decay because the stored memory each new chunk reads was itself built in conditions the model never trained on.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Keep each stored entry in distribution where it is made, by caching every chunk while it attends only to itself.
- 4 · Choose your own research ~3 min
- 5 · How it works ~3 min
- 6 · The experiment ~6 min You’ll guess: Against prior work, how did ID-Forcing do on a standard video generation benchmark and on mitigating drift?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.