In-Distribution Forcing Prevents Drift in Long Video Generation
October 1, 2026
In-Distribution Forcing (ID-Forcing) mitigates color and texture drift in autoregressive video diffusion models by aligning KV caching with the training distribution. This addresses the KV-provenance problem where cached entries become out-of-distribution during long rollouts.
HOW THIS AFFECTS YOU
●
builderYou can generate longer, more stable videos without the typical decay in motion dynamics.
●
researcherThis addresses the fundamental OOD problem in autoregressive KV caching.