diffengine

A Python toolbox for training diffusion models that leans on diffusers and MMEngine, with pre defined configs, a model zoo, and a Colab notebook to start from.

DiffEngine describes itself as the open source toolbox for training state of the art diffusion models, built on diffusers and MMEngine. The supported algorithms list starts with Stable Diffusion, Stable Diffusion XL, and DreamBooth, and widens to include LoRA, ControlNet, DeepFloyd IF, and a long tail of other methods. The README leans on MMEngine for a unified config system and modular design, so hyperparameters and loss functions are set from one place.

Training starts from a pre defined config. You pick one from the configs directory, a DreamBooth config being the example in the README, then run a single command with the selected config. Outputs land in a work directory that mirrors the config name, and once training finishes, inference runs through the diffusers pipeline module by pointing at the saved model path. Before any of this, PyTorch version 2.0 or higher has to be installed following the official guide.

A Colaboratory notebook demonstrates the workflow with SD v1.5 and SD v2.1 DreamBooth configs. The model zoo lays out supported algorithms across the Stable Diffusion, SDXL, and DeepFloyd IF families. The acknowledgements are specific: the repository borrows its architecture and part of its code from mmengine, mmagic, and diffusers, and it thanks Z by HP for sponsoring workstations used in development.

DiffEngine picks a clear lane by building directly on MMEngine and diffusers instead of reimplementing them. The pre defined config approach lowers the barrier for a first training run, and the model zoo gives a sense of scope.

• Source 1 · Project repository
• Source 2 · README