selfdiffusion 0.1.5

Easily create datasets for diffusion models personalization

selfdiffusion

selfdiffusion is a CLI tool that helps you prepare training datasets of portrait images for stable diffusion model fine tuning.

selfdiffusion takes as input images of different sizes and uses an empirical heuristic to extract the subject of the image and create a version of it that fits the native resolution of the model you want to fine-tune.

Getting started

install the CLI tool with pip:

pip install selfdiffusion

Usage

Start by assembling a set of images that contains the single subject you want to fine tune the model on. If you don't have such a dataset, we provide the utiliy bing command to download images from the internet into a target folder:

selfdiffusion bing "portrait of a person" --target /path/to/target/folder --limit 20

this will download 20 images of people from bing and store them in the /path/to/target/folder.

Once you have your starting images, extract portraits suitable for fine tuning with the following command:

# 512x512 is ideal for fine-tuning runwayml/stable-diffusion-v1-5
selfdiffusion person /path/to/input/images /path/to/output/images --resolution 512 512

you can find out more about model fine tuning in the exemples folder of the diffusers repository.

Notes

selfdiffusion will produce an output image for every input image given that:

1. The input image as a resolution that is greater then the requested output resolution in both dimensions.

2. There is a single person in your image, as determined by yolov8 and it's face is the only one in the image, as determined by face_recognition.

Sometimes, selfdiffusion will filter out an image that very evidently contains a single subject and it's face. If you absolutely want to have this image in your dataset, we suggest using a tool like Birme for capturing the subject and resizing to the native model resolution.

Author

Jonathan Pelletier.