Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE
SplitMoE separates video diffusion experts by role instead of forcing uniform token routing.
The paper says conventional visual MoEs fragment coherent video patches by balancing expert usage too aggressively. SplitMoE divides the expert pool into semantic experts and generic experts, then uses prototype-guided routing with pull-push regularization. The authors report better convergence, routing coherence, and video generation quality under the same activated-parameter budget. It was accepted as a NeurIPS 2026 Spotlight paper. ArXiv · AI/CL/LG's note
The paper says conventional visual MoEs fragment coherent video patches by balancing expert usage too aggressively. SplitMoE divides the expert pool into semantic experts and generic experts, then uses prototype-guided routing with pull-push regularization. The authors report better convergence, routing coherence, and video generation quality under the same activated-parameter budget. It was accepted as a NeurIPS 2026 Spotlight paper. ArXiv · AI/CL/LG's note
score 4