Researchers have introduced PARE, a novel method for making Video Diffusion Transformers (DiTs) more computationally efficient. PARE addresses the high compute demands of DiTs by jointly compressing model width and depth through structure-aware pruning and input-adaptive routing. The system intelligently prunes attention heads based on their spatial or temporal roles and employs a lightweight router to dynamically select blocks for execution based on denoising timestep and visual content. Experiments on the Wan2.1-14B dataset for image-to-video and text-to-video generation demonstrate that PARE significantly reduces per-step computation while maintaining video quality. AI
IMPACT This research offers a method to reduce the computational cost of video generation models, potentially enabling wider adoption and faster iteration.
RANK_REASON The cluster contains a research paper detailing a new method for improving AI model efficiency.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
