The authors analyzed autoregressive video diffusion models and identified that drift arises from the KV-provenance problem: beyond the training horizon, KV entries are cached under unseen provenance, becoming OOD.
Test-time method prevents drift in long video generation by keeping caching in-distribution
ID-Forcing aligns both KV caching and conditioning with training configurations, enabling minute-scale generation from a five-second model.
Top University
Jeongwoo Shin · Youngyoon Choi · Sangwoo Jo · Hyunmog Kim · Sungjoon Choi · Joonseok Lee · +2 more
Seoul National University · Korea University · Sungkyunkwan University · Georgia Institute of Technology
Research Digest··2 min read
Thread:KV Cache Compression
The paper identifies the KV-provenance problem, where cached key-value entries become out-of-distribution beyond the training horizon, causing drift in autoregressive video generation.
Why this paper
From Seoul National University and 3 others · Part of KV Cache Compression, now 4 papers
In one line
In-Distribution Forcing prevents drift in long video generation by aligning KV caching and conditioning with training configurations.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§