SparkDiffusion

A unified post-training framework — compensated sparse attention, trajectory-mixed distillation, and fused FP8 deployment — that turns dense video DiTs into high-sparsity, few-step generators with up to 265× measured end-to-end speedup on a single GPU.