AI Image Glossary and Quick Reference

This is an independent educational reference on jsrkalonda.com.

This page defines the words you meet when you generate, edit, and publish AI images. Terms are grouped by workflow stage: generation, prompting, and rights and provenance. Each definition runs two to four sentences and includes one concrete detail you can check in your own tool, such as where a seed value appears or what a checkpoint file contains. A final section covers accessibility and publishing basics, including a short checklist and FAQ. Use it as a lookup while you work, not as legal advice.

1. How to use this glossary

Terms are grouped by the stage of the workflow where they usually come up. Core generation terms cover the machinery behind an image; prompt-writing terms cover how you steer it; rights and provenance terms cover what you can claim and how you record it. Read the section that matches your current task rather than the whole page.

Each definition is two to four sentences and carries one concrete detail: a file extension, a parameter name, a field you can inspect, or a one-line example. When a term has a common variant spelling, the entry uses the most widespread form. Where a term is defined by an external body, the entry links to that body at first mention.

Definitions describe how tools commonly behave, not how every tool behaves. Interfaces change; the underlying concepts are more stable. If a definition conflicts with your software's documentation, trust the documentation for that specific version.

2. Core generation terms

Prompt: The text instruction that describes the image you want, such as "a red bicycle leaning on a brick wall, overcast light." Most interfaces pass it to the model unchanged, so word order and specificity matter.
Negative prompt: A separate text field listing what to avoid, such as "blurry, extra fingers, watermark." Many tools apply it as a steering signal away from those concepts rather than a hard filter.
Seed: A number that initializes the random noise a generation starts from. Reusing the same seed with the same prompt and settings usually reproduces a near-identical image, which makes seeds useful for comparison.
Latent diffusion: The family of methods that denoise an image in a compressed latent space instead of full pixels. Working in latent space is why a 1024-pixel image can be generated on consumer hardware.
Model weights: The stored numeric parameters a model learned during training; they are the bulk of a model file's size. Weights are usually distributed as files of several gigabytes with extensions like .safetensors or .ckpt.
Checkpoint: A single file holding a complete set of weights for a model, ready to load. A checkpoint is the whole model, whereas a LoRA is a small add-on applied to one.
Fine-tune: Further training of an existing model on a narrower dataset, such as a specific illustration style. A fine-tune changes the weights themselves, so it produces a new checkpoint.
LoRA: Low-Rank Adaptation: a small trained file that adjusts a base model's behavior without replacing it. LoRA files are often tens to a few hundred megabytes and are loaded alongside a checkpoint.
Inference: The act of running a trained model to produce output, as opposed to training it. Each generated image is one inference pass, and step count and resolution drive how long it takes.
Upscaling: Increasing an image's pixel dimensions, either by interpolation or by a model trained to add plausible detail. A 512-pixel generation upscaled to 2048 pixels is a common workflow.
Inpainting: Regenerating a masked region of an existing image while leaving the rest intact. You paint a mask over an object and describe what should replace it.
Outpainting: Extending an image beyond its original borders by generating new content around the edges. It uses the same masking idea as inpainting, applied to the outside of the frame.
Model card: A short document describing a model: its intended use, training data summary, and known limitations. Hugging Face popularized the format by showing a model card on each hosted model page.

3. Prompt-writing terms

Subject: The main thing in the image, stated first and plainly, such as "a ceramic teapot." Putting the subject before modifiers keeps it from being diluted by later words.
Composition: How elements are arranged in the frame: close-up, wide shot, centered, rule of thirds. Naming a camera framing like "medium shot" often changes layout more than adjectives do.
Style anchor: A short phrase that fixes the visual register, such as "flat vector illustration" or "35mm film photograph." One clear anchor usually beats a long list of style adjectives.
Aspect ratio: The width-to-height proportion of the output, written as 16:9, 1:1, or 4:5. Many tools set it through a resolution parameter or a dedicated ratio control.
Token weighting: Syntax that increases or decreases a word's influence, often written as (term:1.3). It is a nudge, not a guarantee, and extreme values tend to distort results.
Prompt log: A saved record of the exact prompt, negative prompt, seed, model, and settings used for an image. Keeping one lets you reproduce or audit a result later.
Variation: A new output generated from the same prompt with a different seed. Variations share a prompt but differ in layout and detail.
Remix: Reusing an existing image's prompt or structure as the starting point for a new generation. Remixing a prompt is fine; remixing a protected image raises the rights questions covered in section 4.

4. Rights and provenance terms

Human authorship: The requirement that a copyright claim rest on creative choices made by a person. The U.S. Copyright Office AI initiative publishes guidance on this point; purely machine-generated output is not registrable on its own.
Training data: The collection of images and text a model learned from. Model cards and dataset pages are the usual places to look for a summary of what was included.
Likeness: A real person's recognizable appearance. Generating a recognizable individual's face can raise publicity and personality-rights issues separate from copyright.
Trademark: A brand name, logo, or other sign identifying a commercial source. A generated image that reproduces a logo may infringe even if the artwork itself is new.
Content provenance: Recorded information about how a piece of media was made and changed. It answers "where did this come from" rather than "is this good."
C2PA: The Coalition for Content Provenance and Authenticity, which publishes an open technical standard for attaching provenance data to media. See C2PA for the specification.
Content Credentials: A provenance label built on the C2PA standard, implemented in tools such as Adobe Firefly. It can record that an image was AI-generated and which tool made it.
Synthetic content: Media generated or substantially altered by AI. Labeling synthetic content as such is increasingly expected by platforms and is good practice regardless.

5. Accessibility and publishing basics

Every published AI image still needs alt text. Screen-reader users get the same information from a page only if the image has a text equivalent, and AI images are no exception. For guidance on writing concise, useful descriptions, see the WebAIM alt text guide.

Use this checklist before an AI image goes live:

Alt text: describe the image's content and purpose in one or two sentences; use an empty alt="" only for purely decorative images.
File size: compress and resize to the dimensions actually displayed; a 4000-pixel file in a 600-pixel slot wastes bandwidth.
Honest caption: state that the image was AI-generated when that is relevant to the reader.
Provenance record: keep the prompt, seed, model, and any Content Credentials attached to the file.
License check: confirm the model's license and that no protected logo, trademark, or recognizable likeness is reproduced.

Does an AI image need alt text if it is decorative?

If the image adds no information, mark it decorative with an empty alt attribute so screen readers skip it. If it carries meaning, write alt text as you would for a photograph.

Can I copyright an image I generated?

Copyright protects human creative expression. Under current U.S. Copyright Office AI initiative guidance, purely machine-generated output is not registrable, while human contributions such as selection and arrangement may be. Consult a lawyer for your situation.

How do I show that an image is AI-generated?

Add a visible caption and, where supported, keep Content Credentials from a tool like Adobe Firefly. Both together are stronger than either alone.

Where do I find a model's intended use?

Check its model card, a format popularized by Hugging Face, which typically lists intended use, training data summary, and limitations.

For the broader workflow, return to the main guide; for prompt technique, read the step-by-step method; and for licensing and attribution questions, see the copyright and provenance guide.