Autoregressive (AR) video diffusion models share an old ailment: as clips get longer, colors and textures slowly shift and motion decays — the field calls it drifting. The dominant fix is KV conditioning, which selects or rewrites cached key-value entries at inference to control what the model reads. A four-school team from Seoul National University, Korea University, Sungkyunkwan University and Georgia Tech offers a different diagnosis: managing what is read is not enough, because the failure starts at how entries are written. The paper is on arXiv (https://arxiv.org/abs/2610.03120), with code released alongside.

KV provenance: entries go out-of-distribution at write time

The observation is blunt: KV conditioning implicitly assumes cached entries stay in-distribution, and the assumption breaks beyond the training horizon. A KV entry depends on the entries it attended to when it was cached — its provenance. Trained only on short horizons, the model writes every new chunk during rollout under a provenance never executed in training, so the entries themselves become out-of-distribution (OOD). The team names this the KV-provenance problem, with a harsh corollary: provenance is fixed at caching time, and no conditioning policy can recover from it.

Two rules: self-caching on write, an exact window on read

ID-Forcing is a pure test-time framework — no fine-tuning — with one rule per KV operation:

  • Write side, self-caching: the earliest entries in the window are cached attending only to themselves, the rest autoregressively on top. Every entry is written under a provenance training actually executed, preventing OOD at the source.
  • Read side, exact rolling window: the first chunk k0 — which every training step conditions on — is permanently kept, alongside the most recent entries, re-rotated so temporal distances stay in the trained range, with window length strictly L <= N-1.

Results: a 48x training-free stretch

On the project page, videos from a single prompt are extrapolated to 30/60/120/240 seconds from a roughly 5-second training horizon, with no fine-tuning. The repo's default scripts generate 2-minute videos — 160 chunks, 1917 frames at 16 fps. Evaluation spans two backbones: ID-Forcing takes over Self-Forcing at chunk 7 (the first chunk past its 7-chunk training horizon) and LongLive at chunk 1 (itself trained on minute-long videos). The method stays competitive on standard video-generation benchmarks while substantially outperforming prior work on drift, validated by two self-built drift metrics and a user study: color drift compares 180-bin HSV hue histograms of the first and last frames; motion drift uses VBench Dynamic Degree over the first and last 5 seconds (RAFT optical flow, 8 fps subsampling). Baselines include Self-Forcing, Deep Forcing, MemRoPE and Infinity-RoPE.

Open source and reproduction

The GitHub repo is clean: two one-shot scripts drive Self-Forcing and LongLive; the checkpoint table specifies EMA weights for Self-Forcing, base+LoRA for LongLive, with the text encoder and VAE from Wan2.1-T2V-1.3B; evaluation uses 128 MovieGen prompts; code keeps the Apache License 2.0 of both base repos. Interrupted runs resume out of the box — existing videos are skipped.

So what

The value is not "longer videos again" but the cost structure: minute-scale horizons demand minute-scale data and compute, while ID-Forcing shows that aligning KV reads and writes with the training distribution at test time lets an off-the-shelf 5-second model extrapolate stably. For compute-limited teams it is another case of trading inference-time engineering for training-time spend — and a reminder that the KV cache is not just a memory-saving artifact but the very carrier of a distribution problem. The hot cache research on the text side and this thread on the video side are telling the same story.