Live page · Day archive

MC-Sparse: Deconstructing and Closing the Dense-Sparse Attention Gap in Diffusion Transformers

Read the original at HF Daily Papers

Summary

Researchers introduce MC-Sparse, a training-free framework that improves diffusion transformer efficiency by caching metadata and selecting key-value tokens to maintain output fidelity during video and 3D generation.

Carried by: HF Daily Papers. First seen: .