In-Context Forcing: New video diffusion method enables parallel denoising, beats SOTA on VBench
Progressive noise contexts fix temporal consistency and unlock parallel frame generation for 2x faster inference.
Autoregressive video diffusion models typically generate frames sequentially, conditioning each new frame on previously fully denoised, clean frames. According to the new paper "In-Context Forcing: Uncovering Context Effects in Autoregressive Video Diffusion," these clean frames leak excessive local details, causing the model to take shortcuts and compromise temporal semantics and dynamics. The researchers instead explore the impact of noisy contexts, showing that simply applying same-level noise provides insufficient guidance and leads to poor temporal consistency.
The proposed In-Context Forcing paradigm uses contexts with decreasing noise levels—applying less masking to distant frames and more masking to adjacent ones. This adaptive guidance ensures both robust temporal consistency and high inter-frame dynamics. Key to its impact, the method decouples strict dependence on previous clean frames, enabling cross-frame parallel denoising. This achieves substantial inference acceleration without sacrificing performance. Experiments on VBench demonstrate that the method significantly outperforms state-of-the-art approaches in both visual fidelity and inference speed, marking a step forward for efficient high-quality video generation.
- Uses decreasing noise levels for context frames: less masking for distant frames, more for adjacent ones
- Enables cross-frame parallel denoising, decoupling from clean-frame dependence for major inference speedups
- Outperforms state-of-the-art methods on VBench across visual fidelity and inference speed metrics
Why It Matters
Faster, more coherent video generation accelerates production pipelines for creators, researchers, and real-time applications.