SparkDiffusion
Unified framework for up to 265x single-GPU acceleration of visual generation.
SparkDiffusion identifies the high-sparsity trap in video diffusion transformers — at extreme attention sparsity, step-local training losses keep decreasing while terminal generation quality stagnates or degrades — and turns the diagnosis into a staging principle: first adapt the sparse architecture into a coarse prior, then correct the terminal distribution.
The framework combines a short sparse warm-up, few-step trajectory-mixed distillation, and FP8 quantization with fused kernels. It sustains 97% attention sparsity with strong visual quality on long-sequence 720P generation across Wan2.1/Wan2.2 backbones and T2V/I2V tasks, achieves a 265x end-to-end speedup over the 50-step CFG dense baseline on a single RTX 5090 (220x on H100) with 3-step CFG-free inference, and denoises a Wan2.1-T2V-1.3B-480P video in 1.3s.