Stable Fast is an open-source inference optimization framework that squeezes maximum performance out of Hugging Face Diffusers models on NVIDIA GPUs. Hosted on GitHub with about 1,300 stars under the MIT license, it compiles diffusion pipelines into highly optimized execution graphs, with code updates through 2025.
Diffusion models are slow to run, and existing accelerators such as TensorRT or AITemplate can take dozens of minutes to convert a model. Stable Fast instead uses torch.compile and CUDA graphs to optimize pipelines quickly, combining operator fusion, memory optimization and dynamic shape handling to reach state-of-the-art inference speeds across all kinds of diffuser models โ including image models and Stable Video Diffusion. It works with the standard Diffusers API, so existing pipelines can usually be accelerated with a few lines of code rather than a rewrite. Prebuilt wheels and a Colab notebook make it easy to benchmark the speedup on your own model before adopting it. The author notes that active development has paused while a successor project (Comfy-WaveSpeed, aimed at ComfyUI) is built, and the same technology powers hosted, pre-optimized models on WaveSpeed.ai.
The Stable Fast library is free and open source under the MIT license, so self-hosting costs nothing beyond your GPU. A related hosted service, WaveSpeed.ai, offers pre-optimized models such as FLUX and Wan with LoRA support as a paid cloud option for those who prefer not to run their own infrastructure.
ML engineers and product teams serving diffusion models in production who need lower latency and higher throughput on NVIDIA hardware without rewriting their Diffusers pipelines.