The Art of Consistency — Episode 2
Introduction
What follows is a continuation of The Art of Consistency. An exploration of using Stable Diffusion, LoRAs, and CustomGPTs to create consistent characters. This is the second article in this series. At the end, I will provide links to all referenced tools and resources so you can reproduce this experiment yourself.
The inspiration image comes from a Python script that randomly generates prompts from over 20 dictionaries of physical and aesthetic characteristics, each containing a list of many different variations. I have generated a large number of images using a LoRA named PeopleProject, created from a dataset of 1,000 images.
From the new images created from PeopleProject, I select someone that looks interesting and pass the image to VisionTaggerGPT. This CustomGPT processes an image and generates a structured JSON file containing all the necessary data for the experiment.
The most important part of the JSON provided by VisionTaggerGPT is the physical tags. These are the beginning of every prompt and the one item that is held constant throughout this experiment. Below are images generated solely from the physical tags, serving as the foundation for the variations explored in this experiment.
The following sections will demonstrate how this LoRA preserves character consistency while allowing variations in environment, lighting, and style. It will highlight the usefulness of VisionTaggerGPT as a useful tool in creating meaningful data, and hopefully show how consistency in character creation is possible through the use of Stable Diffusion when done in a systematic way.
---------------------------------------------------------------------------------------------------------