A research team has found a way to make large video generation models run 30 times faster without rebuilding them from scratch.
Video diffusion models produce high-quality output but chew through compute at every step of the process. Existing shortcuts, called few-step distillation, reduce the number of denoising passes but treat the model as a fixed structure throughout — wasting capacity at stages that don't need it. The new framework, Dynamic-in-Few-Step, attacks that waste directly: it sparsifies the model's architecture differently at each denoising timestep, then bakes those structural decisions into a distillation process so everything trains together. The result is what the authors call a Mixture-of-Models — a set of step-specific, leaner subnetworks rather than one monolithic network running at full cost regardless of context.
The numbers are notable. On Wan-14B, a 14-billion-parameter video model, the method strips out 24% of per-step floating-point operations on top of an already accelerated 4-step distillation run, adding a further 1.2x wall-clock gain. That compounds to a 30x speedup over the original 50-step baseline while keeping generation quality competitive. Crucially, the authors say the technique stacks on top of other acceleration methods rather than replacing them.
Video generation is fast becoming the most compute-hungry corner of the AI industry, and labs are under real pressure to make models cheaper to run, not just cheaper to train. A 30x inference gain on a model this size would meaningfully change the economics — though "competitive quality" is doing some heavy lifting until independent evaluations weigh in.