AI & Design Glossary

Plain-language definitions of the AI terms that show up in design work, written for what they mean at your desk, not for the maths underneath.

Every entry here is defined through a design lens: what the term changes about your work, when you would reach for it, and where to see it in practice. Terms are grouped by the part of the process they belong to and listed alphabetically within each group. Where a term carries a legal or ethical weight, the definition says so plainly rather than waving it away.

Image Generation

CFG scale

How strictly the model obeys your prompt. Push it up and the model sticks closely to what you asked, sometimes stiffly. Pull it down and it takes more liberties. When an image ignores half your brief, this is usually the dial to reach for before you rewrite the prompt.

See in practice: Image Generation Tools · Brief Writing with AI · Build Taste, Generate, Refine

ControlNet

An add-on that lets you steer a generation with a structural guide: a sketch, a pose, an edge map, a depth map. It is how you keep a composition you have decided on while letting the model handle surface and style. Structure from you, rendering from the model.

See in practice: Image Generation Tools · Composition and AI · Build Taste, Generate, Refine

diffusion model

The kind of model behind Midjourney, DALL-E, and Stable Diffusion. It starts from visual noise and clears it away step by step until an image appears. You do not need the maths, but it helps to know the output is grown, not retrieved: there is no stock photo underneath, which is exactly why your prompt and references matter so much.

See in practice: Image Generation Tools · The Design Advantage · Build Taste, Generate, Refine

img2img

Feeding the model an existing image and asking it to reimagine that image in a new style, medium, or treatment, rather than starting from text alone. A fast way to carry a composition you like into a different visual register.

See in practice: Image Generation Tools · Moodboarding with AI · Build Taste, Generate, Refine

inpainting

Redrawing one chosen part of an image while leaving the rest untouched. You mask a region and regenerate only inside it. This is the workhorse of refinement: fix a hand, swap a background object, or correct a detail without losing the parts that already work.

See in practice: Image Generation Tools · Build Taste, Generate, Refine · From Prompt to Portfolio

latent space

The internal map a model uses to organise visual ideas before it draws anything. Concepts that look related to the model sit near each other there. The practical takeaway: small wording changes can move you a long way across that map, which is why two near-identical prompts sometimes return wildly different images.

See in practice: Image Generation Tools · Build Taste, Generate, Refine · The Design Advantage

LoRA

A small add-on that teaches a model one specific look, character, or style without retraining the whole thing. For a designer it is the closest tool to a reusable brand filter: train it once on a consistent set of references and apply that visual signature across many generations.

See in practice: Image Generation Tools · Moodboarding with AI · Colour Theory and AI

outpainting

Extending an image past its original edges, as if panning the camera to reveal more of the scene. Useful when a composition is strong but cropped too tight, or when you need a wider format for layout without regenerating the whole piece.

See in practice: Image Generation Tools · Composition and AI · Build Taste, Generate, Refine

prompt weight

Syntax that tells the model which words to care about most. Increasing the weight on a word pushes the model to honour it; decreasing it lets the word fade. A precise tool for when an image is right except for one element the model keeps underplaying.

See in practice: Image Generation Tools · Brief Writing with AI · Visual Hierarchy and AI

sampler

The method a model uses to clear away noise as it generates. Different samplers carry slightly different aesthetic fingerprints, some cleaner, some grainier. You rarely change it early on, but it is worth knowing it exists when two runs of the same prompt feel different for no obvious reason.

See in practice: Image Generation Tools · Build Taste, Generate, Refine

seed

The starting number that fixes a generation. Reuse the same seed with the same prompt and you get the same image back. This is your handle on consistency: lock a seed when you have something close and want to change one thing at a time instead of rerolling into a different picture.

See in practice: Image Generation Tools · Build Taste, Generate, Refine · From Prompt to Portfolio

steps

How many passes the model takes to refine an image. More passes add detail up to a point, then stop helping and just cost time. Past roughly twenty to fifty you are usually paying for patience, not quality, so treat it as a tradeoff, not a quality slider.

See in practice: Image Generation Tools · Build Taste, Generate, Refine

upscaling

Enlarging a low-resolution image while keeping its detail clean, often adding texture the original lacked. It belongs at the end of a workflow, once you have settled on a final composition, not while you are still exploring, because the cost adds up fast on images you will discard.

See in practice: Image Generation Tools · From Prompt to Portfolio · Research & Ideation Tools

Language Models & Prompting

agent

A model that does more than chat: it plans a few steps and takes actions, like searching, building, or assembling a result, to reach a goal. Increasingly common in layout and prototyping tools that turn a prompt into a working draft you then direct.

See in practice: Layout & Prototyping Tools · From Prompt to Portfolio · Text & Copy Tools

chain-of-thought

Asking a model to reason step by step before it answers, rather than jumping straight to a conclusion. For design critique and brief analysis this reliably produces more specific, more useful output than a one-line request.

See in practice: Design Critique with AI · Text & Copy Tools · Brief Writing with AI

completion

The text a model generates to continue from your prompt. Give it the opening of a sentence or a brief and it carries on. The most basic shape of working with a language model, underneath every chat interface.

See in practice: Text & Copy Tools · Brief Writing with AI · How to Use ChatGPT for Design

context window

How much conversation and material a model can hold in mind at once, measured in tokens. Go past it and the earliest details quietly drop. For design work this is why a long, rambling chat starts forgetting your brief: feed it the essentials, not the whole history.

See in practice: Text & Copy Tools · Brief Writing with AI · How to Use ChatGPT for Design

embedding

Turning text or images into a string of numbers that captures their meaning, so a machine can judge what is similar to what. The quiet engine behind visual search and reference recommendation, where you find images by feel rather than by keyword.

See in practice: Research & Ideation Tools · Moodboarding with AI · Image Generation Tools

few-shot

Teaching by example. Show the model two to five worked input-output pairs before asking for a new one, and its results get far more consistent. The fastest way to pin down a format or tone without writing a paragraph of instructions.

See in practice: Text & Copy Tools · Brief Writing with AI · Design Critique with AI

fine-tuning

Training a model further on your own material so it specialises in your voice or domain. It is slower and more costly than prompting, and for most design work a good brief and a few examples get you most of the way without it.

See in practice: Text & Copy Tools · Brief Writing with AI · The Design Advantage

hallucination

When a model states something false with complete confidence. It is not lying, it is filling a gap with a plausible guess. The confident-looking errors are the dangerous ones, which is why a working eye and a fact-check matter most exactly where the output looks most finished.

See in practice: When AI Gets It Wrong · Design Critique with AI · Text & Copy Tools

RAG

Retrieval-Augmented Generation: handing a model the real documents it should answer from, so its reply is grounded in your material rather than its training. The reliable way to get answers about a specific client, brand, or brief instead of a generic best guess.

See in practice: Text & Copy Tools · Brief Writing with AI · Research & Ideation Tools

system prompt

The standing instruction that sets a model’s role and limits before your conversation begins, like casting direction before a shoot. In most consumer tools you never see it, but in tools that let you set one it is the single most effective place to fix tone and behaviour.

See in practice: Text & Copy Tools · Brief Writing with AI · Design Critique with AI

temperature

A setting that governs how predictable a model’s output is. Low keeps it tight and repeatable; high makes it loose and surprising. Reach for low when you want a reliable format, high when you are fishing for unexpected directions early in a project.

See in practice: Text & Copy Tools · Build Taste, Generate, Refine · Research & Ideation Tools

token

The small chunk a model breaks text into before it reads it. A word can be several tokens. You care about this for two practical reasons: it sets what a model costs to run, and it sets how much it can hold in mind at once.

See in practice: Text & Copy Tools · Brief Writing with AI · How to Use ChatGPT for Design

tool use

When a model reaches outside the conversation to call a function, an API, or a design tool, deciding for itself when to do so. It is what lets an assistant fetch live information or build something rather than only describe it.

See in practice: Layout & Prototyping Tools · Text & Copy Tools · Client Disclosure

Design & AI Workflow

AI moodboard

A reference collection of images, colours, and directions assembled quickly with AI to capture a mood before you commit. It speeds up the part of the process where you decide what good looks like, which is the part that most shapes the final work.

See in practice: Moodboarding with AI · Build Taste, Generate, Refine · Research & Ideation Tools

AI-assisted mockup

Building a draft layout yourself, then letting the model fill backgrounds, draft copy, or stand in placeholder images. A hybrid workflow where the structure and decisions stay yours and the model handles the filling-in.

See in practice: Layout & Prototyping Tools · From Prompt to Portfolio · Build Taste, Generate, Refine

aspect ratio

The proportions of your final image, such as 16:9, 1:1, or 9:16. Decide it before you generate, not after, or you will waste runs cropping into compositions the model never framed for that shape.

See in practice: Composition and AI · Build Taste, Generate, Refine · Image Generation Tools

brand lock

The set of constraints that keep AI-generated work on-brand: fixed logo placement, an allowed colour range, type rules. Without them, generation drifts off-brand fast; with them, you can move quickly and still stay consistent.

See in practice: Colour Theory and AI · Build Taste, Generate, Refine · Layout & Prototyping Tools

composition constraint

A structural rule you give the model, such as rule-of-thirds, asymmetrical balance, or a centred focal point. It guides the layout without dictating every pixel, and it is the lever that pulls a generation out of the model’s default centred symmetry.

See in practice: Composition and AI · Visual Hierarchy and AI · Build Taste, Generate, Refine

generative fill

Selecting empty or unwanted space in an image editor and having the model paint in something that fits, whether to extend a background or remove an object. A precise, in-context tool that keeps you inside your working file.

See in practice: Layout & Prototyping Tools · Moodboarding with AI · Image Generation Tools

iteration loop

The cycle of generate, critique, refine, run fast enough that you can explore ten directions in the time it once took to sketch one. The speed is the point only if your critique is good: it multiplies whatever judgement you bring to each pass.

See in practice: Build Taste, Generate, Refine · Design Critique with AI · From Prompt to Portfolio

prompt template

A reusable instruction with blanks you swap out, such as redesign this object in this style for this audience. It saves you rewriting the same scaffolding and, more usefully, makes your prompting repeatable and reviewable.

See in practice: Text & Copy Tools · Brief Writing with AI · Build Taste, Generate, Refine

reference image

A picture you hand the model to anchor what you want, whether for style, composition, or colour. Often worth more than a paragraph of description, because it tells the model what you mean instead of what you can spell out.

See in practice: Moodboarding with AI · Build Taste, Generate, Refine · Image Generation Tools

style transfer

Applying the visual style of one image to the content of another, so a photo can take on a painting’s treatment or a brand’s palette. Useful for testing a direction quickly, with the usual caution about whose style you are borrowing.

See in practice: Moodboarding with AI · Colour Theory and AI · Copyright and Licensing

variation batch

Generating several versions of a design at once, then choosing among them rather than chasing one perfect result. A good fit for early exploration, where seeing four real options teaches you more than imagining one.

See in practice: Build Taste, Generate, Refine · Layout & Prototyping Tools · Moodboarding with AI

vectorising

Converting a pixel image into editable vector paths, so it scales to any size without going soft. AI tools automate the path-drawing that used to be slow and manual, which matters whenever a raster generation needs to become a logo or print-ready mark.

See in practice: Image Generation Tools · Layout & Prototyping Tools · Composition and AI

voice clone

Synthesising audio in a particular person’s voice, used for narration or multilingual voiceover in a consistent tone. Powerful and sensitive in equal measure: using someone’s voice is a consent and disclosure question, not just a technical one.

See in practice: Client Disclosure · Research & Ideation Tools · When AI Gets It Wrong

AI-assisted label

A plain badge, such as partially AI-generated, on work that mixes your craft with AI tools. A growing best practice in portfolios and client work, and a simple way to be transparent without underselling the judgement you brought.

See in practice: Client Disclosure · From Prompt to Portfolio · Copyright and Licensing

attribution

Crediting the sources behind a piece of work. With AI the honest version is twofold: respect the rights in any source material, and be transparent that AI was involved when it matters to the client or audience.

See in practice: Client Disclosure · Copyright and Licensing · When AI Gets It Wrong

C2PA

The Coalition for Content Provenance and Authenticity: a standard for attaching a signed record of who created and edited an image. Adoption is still early, but it is the most likely path to provenance you can actually verify.

See in practice: Copyright and Licensing · Client Disclosure

copyright holder

The person or entity that owns a work’s exclusive rights. For AI output, most jurisdictions vest copyright in a human, but ownership of anything derived from training material remains contested. When the holder is unclear, so is your right to use the work.

See in practice: Copyright and Licensing · Client Disclosure · When AI Gets It Wrong

derivative work

A new creation built on existing copyrighted material. An AI image clearly modelled on a known artist or painting may count, and ownership of such work is far from settled. The four-factor matrix in the copyright article is how you reason about the risk.

See in practice: Copyright and Licensing · Client Disclosure · When AI Gets It Wrong

fair use

A legal doctrine allowing limited reuse of copyrighted work for things like commentary, criticism, or teaching. It is fuzzy in court and especially untested for AI, so treat it as a question for a lawyer, not a licence you can assume.

See in practice: Copyright and Licensing · Client Disclosure · When AI Gets It Wrong

indemnification

A vendor’s promise to cover your legal costs if their tool’s output lands you in an infringement claim. Some AI vendors offer it, many do not, and the difference belongs in your decision about which tool to put near paying client work. Check the terms.

See in practice: Copyright and Licensing · Client Disclosure

model license

The terms that say how you are allowed to use a model: commercial work, research only, or open to modify. They vary widely between tools, and reading them is the difference between a clearance you can defend and one you are assuming.

See in practice: Copyright and Licensing · Image Generation Tools · Client Disclosure

opt-out

A mechanism for keeping your own work out of AI training sets. Registries and tools exist to claim your images, though enforcement is still young. Worth knowing both as a creator protecting your work and as a sign of how a vendor treats artists.

See in practice: Copyright and Licensing · Client Disclosure

provenance

The documented origin and edit history of a piece of work: who made it, when, and how. It underpins copyright claims, and because AI output rarely carries clear provenance, building your own record of decisions becomes part of professional practice.

See in practice: Copyright and Licensing · From Prompt to Portfolio · Client Disclosure

synthetic media disclosure

Labelling content as AI-generated or altered. Platforms increasingly require it, and even where they do not, disclosing builds trust. The legal floor is rising, but the professional case for honesty was always there.

See in practice: Client Disclosure · Copyright and Licensing · When AI Gets It Wrong

training data

The vast set of images and text a model learned from. Because the sources are often unclear, so is the legal footing of what comes out. Vendors increasingly disclose their data or offer an opt-out, and knowing a tool’s data story is part of using it responsibly.

See in practice: Copyright and Licensing · When AI Gets It Wrong · The Design Advantage

watermark

A visible or hidden mark that signals ownership or origin. AI tools often strip ordinary watermarks, which is why cryptographic standards like C2PA are emerging to carry that information in a form that survives editing.

See in practice: Copyright and Licensing · Client Disclosure · When AI Gets It Wrong

Every term here is a tool, and a tool is only as good as the judgement you bring to it. The vocabulary that decides whether an AI output is any good is the same design vocabulary it always was. That is where Design Knowledge & AI starts.

Find us on the Gram

Follow our design inspiration curation (that rhymes) & stay up to date on the latest work from our students

Get Started.

You can enrol any day of the year. We are online and study is self-paced, there is no pressure. Enrol when you are ready to start, from anywhere in the world. If you would like to chat or email, feel free to get in touch.

Brochures, Phone Calls & Questions

You can download a free brochure, book a phone call with one of our course advisors, or simply ask a question.

Other ways to get in touch

Australia 1300 655 485

International +61 1300 655 485

Ask Anything info@thegraphicdesignschool.com

Get a quote accounts@thegraphicdesignschool.com

Acknowledgement of Country
The Graphic Design School acknowledges the Traditional Custodians of Country throughout Australia and their continued spiritual connection to land.
We pay respect to Elders past, present and emerging.
Always was, always will be.
RTO Provider № 91706