arxivcs.DCcs.LG2026-07-13
FlashDiff: Efficient Regional Execution and Scheduling for Diffusion Model Serving
Yaqi Qiao, Ping He, Songrun Xie, Ayush Barik, Chensong Zhang, Zhengzhong Tu, et al.
Diffusion models have become the central backbone for modern image, video, and audio generation, but their efficient service remains a challenge. Unlike autoregressive decoding, diffusion inference repeatedly updates high-dimensional spatial or temporal latents over many denoisin…