This is a fork of the DreamBooth repo with some small changes to make easier to use:
- Fixed a couple of runtime errors I was getting.
- Enabled passing the identifier (the unique prompt word for your subject) in as an argument so you don't have to modify Python code.
- Created a script called
generate_regularization.pyspecifically for making regularization images (see docs below). - Added utilities for pruning the
ckptfile and for turning ackptfile into the Hugging Face diffusers file format. - Tweaked the conda env slightly to accommodate the utility scripts.
- Assumed some defaults in the suggested commands below to make setup easier (things like the location of the
model.ckptfile, regularization images, and training images). - Rewrote the documentation to hopefully make it more widely accessible.
Some people report being able to get this to run on a 3090. I haven't been able to. I've been renting A100s from vast.ai (no affiliation), and I'm able to do a full fine-tune in about 15 minutes. If you use Vast, be sure to bump up your disk space way past the default 10GB. That's not nearly enough. I use 100GB so I never have to think about it.
Note that the workflow described below for generating regularization images can be done on a 3090, but the fine-tuning itself cannot.
If you just installed conda (or you're running on a fresh VM):
% conda init bash
% bash
(If your shell isn't bash, substitute the correct shell command.)
Create the conda environment:
% conda env create -f environment.yaml
If you need to update your conda environment for any reason, use this command:
% conda env update -f environment.yaml
If something went wrong and you want to delete the environment and start again:
% conda env remove -n dream-booth
Activate:
% conda activate dream-booth
Next, move your model.ckpt file into the models directory. This is likely a file downloaded from Hugging Face. Make sure to rename the file to model.ckpt (that simplifies the commands below).
If you want to generate regularization images, you can easily do so (even on a 3090) with the following command:
% python scripts/generate_regularization.py --ckpt models/model.ckpt --prompt "a photo of a dog"
Your 200 images will be placed in the directory inputs/regularization where they need to be for the rest of the process to work.
Once your regularization images are in place, put your training data into inputs/training. Training data should be between 5 - 20 512x512 PNGs of the subject you want to train on.
Next, copy and paste the command below, but make the appropriate changes to --identifier and --class_word:
--identifier: The unique word that you want to use to reference your subject in your prompts. It should be something unlikely to already be in the training data.--class_word: The class of what you're training on. For example: man, woman, person, dog, etc.
python main.py --ckpt models/model.ckpt --name DreamBoothFineTune --gpus 0, --data_root inputs/training --reg_data_root inputs/regularization --identifier UNIQUE_IDENTIFIER --class_word SUBJECT_CLASS
(Optional) When you're done, you can prune the checkpoints to create a smaller ckpt file (credit):
% python scripts/prune.py --ckpt path/to/model.ckpt
(Optional) If you did the training on a remote machine, you can copy the ckpt file to your local machine with a command that looks something like this (replace the port number, IP address, and paths):
% scp -P PORT_NUM root@123.456.789.000:/root/Dreambooth-Stable-Diffusion-Tweaked/logs/training2022-10-25T21-25-03_DreamBoothFineTune/checkpoints/last.ckpt .
(Optional) Turn the ckpt file into the diffusers file format (credit) compatible with the Stable Diffusion Photoshop REST API Server:
% python scripts/convert_original_stable_diffusion_to_diffusers.py --checkpoint_path /path/to/model.ckpt --dump_path /output/path