Autoregressive (AR) video diffusion models have become adept at generating short video clips, but pushing them toward long-horizon generation remains difficult. The central obstacle is drifting: over extended rollouts, colors and textures gradually shift and motion dynamics decay, degrading the coherence of the output.
Prior work has largely addressed drifting through KV conditioning, which selects or modifies cached key-value (KV) entries to stabilize generation. A new paper argues that this approach is insufficient on its own, because it assumes the cached KV entries remain in-distribution — an assumption that breaks down beyond the model's training horizon.
The authors identify what they call the KV-provenance problem: nothing constrains how KV entries are constructed during rollout, so the cached entries themselves can become out-of-distribution (OOD). Once that happens, conditioning on them cannot fully prevent drift.
To address this, the paper proposes In-Distribution Forcing (ID-Forcing), a test-time framework that aligns both KV caching and KV conditioning with the configurations seen during training. Its core mechanism, self-caching, targets the problem at its source by caching each chunk without attending to prior KV entries, keeping the rolling window exactly in-distribution.
According to the authors, ID-Forcing seamlessly extends short-horizon models to minute-scale video generation. Extensive evaluations reportedly keep the method competitive on standard video generation benchmarks while substantially outperforming prior work at mitigating drifting, as measured by the authors' drift metrics and a user study.