The AI Image Model Companion Guide
This blog series follows my attempt to build a rights-conscious AI image workflow from beginning to end. Along the way, you’ll see technical terms that may be unfamiliar if you haven’t spent time reading about AI image generation. Rather than stopping every article to explain the same concepts again and again, I’ve collected them here.
AI Image Model
An AI image model is software that has learned statistical patterns from a very large collection of images and text descriptions. It learns relationships between words, shapes, colors, lighting, composition, and other visual concepts, then uses those patterns to generate new images from a written prompt.
Foundation Model
A foundation model is the large, general-purpose model that already knows what people, animals, buildings, clothing, trees, and thousands of other subjects generally look like. Everything in this project builds on top of a foundation model rather than creating one from scratch.
SDXL
Stable Diffusion XL, usually shortened to SDXL, is the foundation model I originally planned to use for this project. It is widely used and has a large collection of compatible tools and community support.
CommonCanvas-XL
CommonCanvas-XL is another open-source foundation model. What caught my attention is that it was developed with the goal of training on Creative Commons licensed material. Whether it ultimately produces artwork I like remains part of this experiment.
LoRA
LoRA stands for Low-Rank Adaptation. Instead of training an entirely new foundation model, a LoRA teaches an existing model a much narrower skill. In my case, that skill is generating fantasy pen-and-ink illustrations in a specific style. You can think of it as adding a specialty to an artist who already knows how to draw.
Dataset
A dataset is the collection of images used for training. For this experiment, every image in my dataset must have documented permission for AI training under the project’s standards.
Caption
Each training image receives a caption describing what the image contains. The model uses both the image and the caption while learning. Writing good captions turned out to be much more important than I expected.
Metadata
Metadata is information about the dataset rather than the images themselves, such as the image source, license, creator, download location, notes, and file names. Good metadata makes the project easier to document and reproduce later.
Prompt
A prompt is the text you type to tell the model what image you want to generate. For example: “A young bard standing beside a circus wagon, pen-and-ink storybook illustration.”
OneTrainer
OneTrainer is the software I’m using to train my LoRA. It controls the training process and exposes the settings that determine how the model learns.
OneTrainer is the software I’m using to train my LoRA. It controls the training process and exposes the settings that determine how the model learns.
Stability Matrix
Stability Matrix is the launcher and model manager I use to organize my AI image tools. It makes installing and switching between different models much easier than doing everything manually.
Rights-Cleared
Throughout this series, I use “rights-cleared” to describe images whose use for this project can be documented through public-domain status, CC0 dedication, Creative Commons licensing where appropriate, or other clearly documented permission. This is a project standard, not a legal term.
Public Domain
A public-domain work is no longer protected by copyright or was never eligible for copyright protection. Public-domain status depends on the laws of the country involved. This project evaluates it primarily under U.S. law.
CC0 (CC Zero)
CC0, pronounced “CC Zero,” is a Creative Commons tool that allows creators to dedicate their work to the public domain as completely as the law allows.
Creative Commons
Creative Commons provides standardized licenses that allow creators to grant specific permissions for how their work may be used. Different licenses allow different uses, so each image must be reviewed carefully before it is included in the project.
Checkpoint
A checkpoint is a saved file containing the learned information, called weights, of an AI image model. Those weights are the patterns the model learned during training and uses to generate images.
A checkpoint is the main model file loaded into image-generation software. It is different from a LoRA, which is a smaller add-on that teaches or adjusts a specific subject, character, or style.
Dataset
A dataset is the collection of images used for training.
For this experiment, every image in my dataset must have documented permission for AI training under the project's standards.
Dataset
A dataset is the collection of images used for training.
For this experiment, every image in my dataset must have documented permission for AI training under the project's standards.