DSA: Dynamic Step Allocation for Fast Autoregressive Video Generation

Arxiv CS.CV (computer vision)

Proposes Dynamic Step Allocation (DSA) to reduce inference time in autoregressive video diffusion models by dynamically assigning sampling steps per frame. Aims to maintain visual quality while accelerating generation.

text-to-video

Arxiv CS.CV (computer vision)·Thanh-Tung LeYunhan ZhaoMenglei ChaiZhengyang ShenZhe CaoDanhang TangXiaohui XieDeying Kong

7 days ago