SparkDiffusion: Mitigating the High-Sparsity Trap in Video Generation
MELON Research Group
We achieve 265× speedup on a single RTX 5090 for 720P video diffusion, while discovering and mitigating the high-sparsity trap — a failure mode where training loss decreases but generation quality collapses at 97% sparsity.
Research NotesVideo DiffusionSparse AttentionDistillationEfficient Inference