thu-ml/TurboDiffusion

Combines SageAttention and Sparse-Linear Attention (SLA) modules for efficient transformer computation, paired with rCM timestep distillation to reduce sampling steps from dozens to 1–4 without quality loss. Supports both text-to-video and image-to-video generation across Wan2.1 and Wan2.2 models at 480p–720p resolution, with quantized checkpoints optimized for consumer GPUs (RTX 5090) and unquantized variants for H100-class hardware.

3,410 stars and 278 monthly downloads. Actively maintained with 1 commit in the last 30 days. Available on PyPI.