This trainer was developed by the Eden team, you can try our hosted version of the trainer in our app. It's a highly optimized trainer that can be used for both full finetuning and training LoRa modules on top of Stable Diffusion. It uses a single training script and loss module that works for both SDv15 and SDXL!

The outputs of this trainer are fully compatible with ComfyUI and AUTO111, see documentation here.

Training images:

Generated imgs with trained LoRa:

The trainer supports 3 default modes:

style: used for learning the aesthetic style of a collection of images.
face: used for learning a specific face (can be human, character, ...).
object: will learn a specific object or thing featured in the training images.

Setup

Install all dependencies using

pip install -r requirements.txt

then you can simply run:

python main.py train_configs/training_args.json to start a training job.

Adjust the arguments inside training_args.json to setup a custom training job.

You can also run this through Replicate using cog (~docker image):

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

Build the image with cog build
Run a training run with sh cog_test_train.sh
You can also go into the container with cog run /bin/bash

Automatic Checkpoint Evaluation

This script uses CLIP img/txt similarity scores to evaluate how good the LoRa is vs how overfit. Download the aesthetic predictor model checkpoint first from google drive. This should give you a file named: aesthetic_score_best_model.pth (99.2 MB)

gdown 1thEIlXVc8lkULVUBY9Ab45tsOERxkjxns

Once the model is downloaded, you can run the eval script with the following CLI args:

output_folder: this is where the outputs of the model get saved as jpeg files
lora_path: path to your LoRA checkpoint (make sure you edit path_to_your_model_checkpoints to point to the correct folder. It generally ends with something like checkpoint-600 where 600 was the training step)
output_json: save all scores in this json file
config_filename: config file used for training

python3 evaluate.py \
--output_folder eval_images \
--lora_path path_to_your_model_checkpoint  \
--output_json eval_results.json \
--config_filename training_args.json

TODO's

Bugs:

pure textual inversion for SD15 does not seem to work well... (but it works amazingly well for SDXL...) ---> if anyone can figure this one out I'd be forever grateful!

Algo:

Improve some of the chatgpt functionality:
- separate the "gpt_description" / "gpt_segmentation" prompt calls and make them run on a subset of prompts in case there's a lot of imgs / prompts (possibly use img_grids for some gpt4-v calls)
- currently some sub-optimal stuff can happen in preprocess() when there's less than 3 or more than 45 imgs, try to improve this
Test if timesteps = torch.randint() can be improved: look at sdxl training code! (see https://github.com/huggingface/diffusers/blob/main/examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py#L1263, https://arxiv.org/pdf/2206.00364.pdf)
Fix aspect_ratio bucketing in the dataloader (see https://github.com/kohya-ss/sd-scripts)
test if textual inversion training can also happen with prodigy_optimizer
improve data augmentation, eg by adding outpainted, smaller versions of faces / objects

Small, minor tweaks:

preprocess.py: the imgs are first auto-captioned and then cropped, this is not ideal, swap this around!

Bigger improvements:

add stronger token regularization (eg CelebBasis spanning basis):
Add multi-token training
implement perfusion ideas (key locking with superclass): https://research.nvidia.com/labs/par/Perfusion/
implement prompt-aligned: https://prompt-aligned.github.io/

Tuning Experiments:

try-out conditioning noise injection during training to increase robustness
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?
offset noise
AB test Dora vs Lora