Finding Best Training Hyper Parameters / Configuration Is Neither Cheap Nor Easy

Currently doing 6 different Stable Diffusion XL (SDXL) fine tuning (DreamBooth like with 5200 regularization images concept) training. My aim is finding best hyper parameters / configuration for masked training of OneTrainer.

Masked training seriously reduces the overtraining when training dataset is not good enough. However it also makes generated picture body disproportional.

Already done more than 4 trainings and currently 6 trainings are running on 2 L40 GPU.

The GPUs are rented from MassedCompute and we already prepared a …

ai aim beginners concept dataset devops diffusion dreambooth easy good however images onetrainer parameters python regularization sdxl stable diffusion stable diffusion xl training