OneTrainer is far from done. Here is a list of things that I would love to implement, but didn't find the time yet.
- Support for more base models: Only Stable Diffusion 1.5, 2.0, 2.1 and the 1.5 and 2.0 Inpainting models are supported at the moment. The goal would be to support all currently released model types.
- Accelerate support for multi GPU training: This will be a bit more complicated. MGDS, the data loader library, needs to be thread safe first.
Some ideas that need to be implemented in MGDS:
- Tag shuffling
- Tag/Caption dropout
- Support for loading concepts from nested folder structures
- Better caching: Make it possible somehow to add and remove concepts or samples from the cache without rebuilding the whole cache. Maybe by splitting the cache into multiple folders based on concepts.
- Adding to that, make it possible to cache multiple versions of a concept within one epoch. This could be used to increase variations of an underrepresented concept.
- Support for class images and prior loss preservation for each concept