ForcingBooster: Adaptive Patch-wise Denoising for Fast Autoregressive Video Diffusion

ForcingBooster. Adaptive Patch-wise Denoising for Fast Autoregressive Video Diffusion. By Zekai Gu and five others, from the Hong Kong University of Science and Technology, VAST, and one other institution. Published on the eighth of October. Its code repository is worked out: about one hundred and sixty source files, with a readme and a licence.

Autoregressive video diffusion models have become a leading approach to real-time interactive video generation and to video world modeling. When they generate the next chunk of video, existing methods usually spread computation evenly across space. Every patch is updated with the same denoising schedule and the same number of function evaluations. Video content is not that even. Static, simple or recurring regions are often easier to denoise, while other regions need further refinement or more context.

A uniform budget therefore cuts both ways. Keep the step count high and easy patches receive computation they do not need. Lower the fixed step count to save time and the difficult patches can be left insufficiently refined.

ForcingBooster is a general framework that stops treating every patch alike. It allocates denoising steps dynamically to individual patches. Easy patches stop early, which is where the efficiency comes from. Difficult patches receive more refinement steps, which is how quality is protected.

The method is meant to sit on top of what already exists. It is compatible with most existing few-step autoregressive diffusion models, and what it adds is a lightweight routing module that needs only modest training. Its job is the per-patch decision: how many denoising steps each patch gets before it is allowed to stop.

In experiments, ForcingBooster reduces the total number of denoising steps applied across video patches by about fifty percent. Visual quality is reported as comparable to the four-step baseline, and the authors describe the overall effect as accelerated generation with negligible quality degradation.

The saving does not come from denoising faster, but from not denoising where nothing is left to fix. Once a patch is done, it leaves the schedule, and the budget that frees up goes to the regions still in need of work. In the reported results, that is roughly half the patch-level steps at a quality close to the baseline it started from.