batch-ti

Batch training script for textual inversion. Keeping it as simple as possible:

1️⃣ Put files in folders

2️⃣ Edit yaml configuration files

3️⃣ Run the script 🚀

Preprocessing

Expected folder structure

The script expects a specific folder structure for your datasets. You'll need to prepare your datasets folder before you start.

Crop the images to 1:1 ratio
Put each set of images in a folder named with the concept label (should be unique)
[optional] add a .txt file containing the description with the same name as the image
Group the concepts based on the template you wish to use. The subject template is a good start, but you can create as many templates as you like.

install requirements

Should be as easy as:

pip install -r requirements.txt

Configuration

batch_config.yaml

Configure the script in batch_config.yaml.

datasets_path: datasets

# Output path for concepts
base_outpath: concepts

# pretrained model to use
model: runwayml/stable-diffusion-v1-5

# deftault initialization value for concept tokens
# you can specify the init for each concept in the batch in [batch_folder]/inits.yaml
default_init: person
default_description: painting
save_every: 100
max_steps: 3000
lr: 3e-04

Specify which batches and templates you want train on:

batches:
  -
    name: my-batch
    templates:
      - subject
  - 
    name: another-batch
    templates:
      - style

Specify custom templates:

templates:
  subject:
    - a photo of a {}
    - a rendering of a {}
    - a cropped photo of the {}

...

inits.yaml

For each batch override the init values for each concept:

my-style: anime
another-style: sketch

Train

Just run it!

python textual_inversion_batch_training.py

Enjoy a walk in the park and a good night of sleep. This will take ~1 hour or more for every concept.

Safety check

Safety checker is disabled for previews during training, enabling it caused OOMs on my gpu. Enable safety at your own risk 🤣

[Extra] Publish to huggingface

You can publish the concepts on huggingface with:

python publish_concepts.py

Acknowledgments

This script is a modified version of https://github.com/fastai/diffusion-nbs. Publishing script is inspired from sd-concept-library Official training colab. You need to join the sd-concept-library organization in order to publish it there.

Eduardishion/train_multiple_concepts_batch_ti