Open AmericanPresidentJimmyCarter opened 1 year ago
Turns out this is already out of date lol. Apparently the better implementation is here: https://github.com/XavierXiao/Dreambooth-Stable-Diffusion
OK -- looking into this more, the dreambooth method finetunes the actual model and so is not portable. This is fine -- lstein has a good, lower memory implementation of the training script that works with the weights from the checkpoint we use otherwise. Then we can just automate the uploading step to huggingface according to the diffusers script above.
Basically run this notebook but as an executor: https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_textual_inversion_training.ipynb
The concepts could be automatically uploaded to huggingface or piped directly as docarrays to
stable
executors.