Trainer

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. A full guide on training can be found in our docs.

<p align="center"> <strong>Training images:</strong><br> <img src="assets/xander_training_images.jpg" alt="Image 1" style="width:80%;"/> </p> <p align="center"> <strong>Generated imgs with trained LoRa:</strong><br> <img src="assets/xander_generated_images.jpg" alt="Image 2" style="width:80%;"/> </p>

The trainer can be run in 4 different ways:

• as a hosted service on our website
• as a hosted service through replicate
• as a ComfyUI node
• as a standalone python script

Using in ComfyUI:

• Example workflows for how to run the trainer and do inference with it can be found in /ComfyUI_workflows
• Importantly this trainer uses a chatgpt call to cleanup the auto-generated prompts and inject the trainable token, this will only work if you have a .env file containing your OPENAI key in the root of the repo dir that contains a single line: OPENAI_API_KEY=your_key_string Everything will work without this, but results will be better if you set this up, especially for 'face' and 'object' modes.

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.

<p align="center"> <strong>Style training example:</strong><br> <img src="assets/style_training_example.jpg" alt="Image 1" style="width:80%;"/> </p>

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):

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

Full unet finetuning

When running this trainer in native python, you can also perform full unet finetuning using something like (adjust to your needs) python main.py train_configs/full_finetuning_example.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!
• figure out why training is 3x slower through comfyui node versus just running main.py as a python job..?
• Fix aspect_ratio bucketing in the dataloader (see https://github.com/kohya-ss/sd-scripts)

Bigger improvements:

• integrate Flux / SD3
• Add multi-concept training (multiple things represented by multiple tokens, trained into a single LoRa)
• add stronger token regularization (eg CelebBasis spanning basis)
• implement perfusion ideas (key locking with superclass): https://research.nvidia.com/labs/par/Perfusion/
• implement prompt-aligned: https://prompt-aligned.github.io/