TurboDiffusion: 100–200× Acceleration for Video Diffusion Models
4 months ago
- #diffusion-models
- #machine-learning
- #video-generation
- TurboDiffusion accelerates video diffusion models by 100-200 times using SageAttention, SLA, and rCM.
- Supports models like TurboWan2.1 and TurboWan2.2 with resolutions up to 720p.
- Install via pip or compile from source; requires Python 3.9+ and torch 2.7.0+.
- Different checkpoints available for GPUs with varying memory capacities (quantized for RTX 5090, unquantized for H100).
- Includes detailed inference scripts for T2V and I2V models with customizable parameters.
- Training code provided with support for FSDP2, Ulysses CP, and SAC.
- Community contributions welcomed, with ongoing improvements and integrations planned.
- Citations provided for related research papers and software.