New framework cuts video AI generation time by 30x
Video Diffusion Models get a 30x speedup while cutting costs by 24%
A team of researchers led by Yu Cheng has unveiled Dynamic-in-Few-Step, a post-training acceleration framework for video diffusion models that reduces computational overhead by dynamically adjusting model sparsity across denoising stages.
Unlike traditional static compression methods, this approach integrates structural sparsification directly into the distillation process, transforming a pre-trained Video Diffusion Model (VDM) into a compact Mixture-of-Models (MoM) optimized for different noise levels. The method introduces a Progressive Training Strategy and Output Rollout Mechanism to stabilize training, while a specialized inference engine ensures efficient deployment. On the Wan-14B model, it delivers a 30x speedup compared to a 50-step teacher pipeline, all while maintaining competitive generation quality.
- Combines dynamic sparsity with few-step distillation for 30x speedup vs 50-step Video Diffusion Models
- Cuts per-step FLOPs by 24% (on Wan-14B) and adds 1.2x wall-clock performance gains
- Introduces new training mechanisms (Progressive Training + Output Rollout) to stabilize model sparsity
Why It Matters
Video generation just got 30x faster and cheaper, unlocking real-time and scalable AI video workflows.