ForcingBooster Halves Denoising Steps in Autoregressive Video Generation

Researchers have posted a preprint on arXiv describing ForcingBooster, a framework that cuts the total number of denoising steps in autoregressive video diffusion models by roughly 50%. The comparison baseline is a four-step model, and the authors say visual quality stays comparable. The method works with most existing few-step autoregressive diffusion models and requires only a lightweight routing module with a small amount of fine-tuning.

Where the compute goes to waste

Autoregressive video diffusion models are now considered one of the leading approaches to interactive real-time video generation and world modeling. Video is built chunk by chunk, and each new segment is generated based on the previous ones. Within a chunk, denoising does the work: it progressively cleans noise out of the latent representation.

The scheme is the same everywhere. All patches get one denoising schedule and the same number of function evaluations (NFE). The authors point to a weak spot: video content is spatially uneven. Static backgrounds, simple textures and repeating elements are easy to reconstruct. Moving objects, new details and regions that need extra context require more refinement.

A uniform compute budget creates a dilemma. If the number of steps stays fixed, computation on easy patches is wasted. If steps are cut for everything at once, hard regions are left underdeveloped and the image suffers.

How adaptive routing works

ForcingBooster dynamically assigns a number of steps to each patch individually. Simple patches finish denoising early, while hard ones get more refinement iterations. The decision is made by a lightweight routing module (the router), which plugs into an already trained few-step model. The authors do not claim to retrain the generator from scratch, and they describe the router's training as modest.

The logic is close to the idea of conditional computation in language models, where hard tokens get more of the compute "attention" than trivial ones. Here the same principle is applied to the spatial dimension of video.

What the claimed results show

• Reduction in total patch-level denoising steps: about 50%
• Comparison baseline: a four-step base model
• Quality: comparable to the baseline, with degradation described as negligible
• Additional components: a lightweight router and modest training
• Compatibility: most existing few-step autoregressive diffusion models

There are also limitations, and the abstract does not address them. Half the steps at the patch level does not necessarily mean half the latency. Patches with different iteration counts are harder to batch efficiently on a GPU, so the final gain depends on the implementation and hardware. The abstract gives no latency or FPS figures. This is a first-version preprint without peer review, and its conclusions rest on metrics that should be rechecked on independent benchmarks.

What this means for teams

For anyone building interactive generation, neural-network game engines or world models, the approach is interesting because it targets the real-time bottleneck: the cost of a single chunk. Since the method is compatible with common few-step models, it can be tested on an existing stack without changing the architecture.

Three things are worth watching during testing. First, real end-to-end latency, not just the step count. Second, behavior on dynamic scenes with abrupt cuts, where almost all patches are "hard" and the savings shrink. Third, artifacts at the borders between patches that stopped at different steps. If the code and weights are released openly, these checks will show whether the paper's twofold savings carry over to production.