Researchers led by Bingxin Xing introduced Sink‑Sparse Hybrid Attention, a way to replace dense self‑attention in pretrained diffusion transformers without retraining. The technique builds static “sink masks” from a small calibration set and adds a light pooled‑KV branch. On the Wan2.1‑T2V‑1.3B model it speeds up end‑to‑end generation by 1.54×, and on Z‑Image‑Turbo by 1.57×, with quality comparable to the original models.
Why it matters
Faster inference lets developers generate images or video more cheaply, which can lower costs for apps that rely on text‑to‑media generation.
