In-Context Forcing: Uncovering Context Effects in Autoregressive Video Diffusion
Published in arxiv , 2026
Lingxiao Yang, Liu Liu, Moran Li*, Han Feng, Wenjian Cao, Jiangning Zhang, Ye Shi
- Equal contribution.
In-Context Forcing introduces a progressive autoregressive paradigm for few-step video diffusion. It conditions each frame on preceding frames with decreasing noise levels, reducing local-detail leakage while preserving temporal consistency and motion dynamics.
The method further introduces a step-wise rolling KV cache for train-test-consistent training and cross-frame causal attention for parallel denoising, substantially accelerating inference without sacrificing generation quality.