Trainer

Code for finetuning and training LoRa modules on top of Stable Diffusion.

Setup

1. Install Replicate 'cog':

sudo curl -o /usr/local/bin/cog -L "https://github.com/replicate/cog/releases/latest/download/cog_$(uname -s)_$(uname -m)"
sudo chmod +x /usr/local/bin/cog

2. Build the image with sudo cog build
3. Run a training run with sudo sh test_train.sh

TODO's

Code / Cleanup:

• make a clean train.py entrypoint that can be run as a normal python command (instead of having to use cog)
• integrate PEFT (https://github.com/huggingface/peft) instead of the hacky, ad-hoc lora stuff
• Modularize the logic in train.py as much as possible, trying to minimize dev work that needs to happen when SD3 drops

Algo:

• add & test proper gradient accumulation so we can train with smaller bs if needed
• Improve the img captioning by swapping BLIP for cogVLM: https://github.com/THUDM/CogVLM
• Add aspect_ratio bucketing into the dataloader so we can train on non-square images (take this from https://github.com/kohya-ss/sd-scripts)
• Implement / Tryout DoRa: https://github.com/catid/dora/tree/main
• right now it looks like the diffusion model gets partially "destroyed" in the beginning of training (outputs from steps 100-200 look terrible), but it then recovers. Can we avoid this collapse? Is the learning rate too high?
• test if textual inversion training can also happen with prodigy_optimizer
• the random initialization of the token embeddings has a relatively large impact on the final outcome, there are prob ways to reduce this random variance, eg CLIP_similarity pretraining.

Bigger improvements:

• Add multi-token training
• implement perfusion: https://research.nvidia.com/labs/par/Perfusion/
• implement prompt-aligned: https://prompt-aligned.github.io/
• make compatible with ziplora: https://ziplora.github.io/