TurboDiffusion

Why we included this project

Waiting minutes for a diffusion model to render seconds of footage is the usual bottleneck in video generation. TurboDiffusion attacks that directly: it combines SageAttention, sparse linear attention, and timestep distillation to claim 100-200x faster end-to-end generation on a single RTX 5090, bringing prompt iteration closer to real time. The repo ships quantized and unquantized checkpoints for several Wan models at 480p and 720p, plus inference scripts, so you can test the speedup on your own hardware. Two caveats: the models were trained on long English prompts, so short or non-English text may need augmentation, and the weights are an evolving research release rather than a finalized product.