CoReDiT: Spatial Coherence-Guided Token Pruning and Reconstruction for Efficient Diffusion Transformers

CoReDiT introduces structured token pruning for Diffusion Transformers using linear-time spatial coherence scoring to identify redundant tokens in self-attention, with coherence-guided reconstruction to maintain dense representations.