Diffusers

Hugging Face maintains this Python package as the common toolkit for running and training diffusion models in PyTorch. It is organized around three swappable parts: pipelines that wrap a full generation workflow, noise schedulers that trade speed against quality, and model classes such as UNets, diffusion transformers and VAEs that recombine into custom systems. Supported families include Stable Diffusion 1.5, SDXL and SD3, Flux, Qwen-Image, plus video models like Wan, LTX and CogVideoX, and audio work such as ACE-Step. Beyond inference it loads LoRA, IP-Adapter and ControlNet weights, offers quantization through bitsandbytes, GGUF, torchao and quanto, supports CPU offloading and torch.compile, and ships example scripts for DreamBooth, LoRA, textual inversion and ControlNet training. Install with pip install diffusers[torch]; an NVIDIA GPU is the practical target, with guides for Apple MPS, Intel Gaudi and AWS Neuron. Releases arrive often, with v0.41.0 out in October 2026. The library is Apache 2.0, though each checkpoint you load keeps its own license. Aimed at developers who generate media from code rather than a GUI.