SplitMoE Architecture for Scaling Video Diffusion Models
September 28, 2026
SplitMoE addresses the uniformity trap in Mixture-of-Experts (MoE) by using a split-role sparse architecture for video data. It prevents routing fragmentation by allowing experts to specialize in semantically distinct spatiotemporal patches.
HOW THIS AFFECTS YOU
●
builderThis architecture offers a path toward more efficient, large-scale video generation models.
●
researcherThis proposes a way to scale visual MoEs without the semantic distortion caused by uniform routing.