4 upvotes
AB-Attention speeds video diffusion transformers without losing quality
What is AB-Attention: Accurate Block Attention for Accelerating Video Diffusion Transformers?
AB-Attention replaces full 3-D attention in video diffusion models with a block-sparse scheme that selects only the most relevant key/value blocks. It keeps the exact attention calculation within those blocks, delivering near-original fidelity while cutting inference cost.
Who it is for
Researchers and engineers building video generators can train faster and run models on cheaper hardware.
How to try it
Visit https://huggingface.co/papers/2610.39663 for code and examples