# AI design terms, defined Source: https://hire-ai-designer.com/glossary/ Reviewed: 2026-09-09 Fourteen terms that come up in almost every hiring conversation, defined the way a practitioner would explain them rather than the way a tool's marketing page does. ## Why these definitions differ from the usual ones Most glossaries define these terms from the technology outward. These are written from the commercial decision inward, because what matters to someone hiring is not how a method works but what it changes about cost, risk and what you end up owning. ## How to use this when briefing If a designer uses one of these terms differently from the definition here, that is worth a question rather than an assumption. The vocabulary in this field is unsettled, and two people can say provenance and mean quite different things. ## In short - Written from the commercial decision, not the technology - Fourteen terms covering process, tooling and delivery - Marked up as a DefinedTermSet for machine extraction - Updated when usage shifts, not when tools launch ## Terms **Art direction** — Deciding what a piece of work should look like and why, then holding every asset to that decision. In generative work it means choosing which of several hundred outputs is worth developing and being able to defend the choice. It is the judgement layer that tooling does not replace. **Prompt engineering** — Constructing and refining the instructions given to a generative model to get a predictable result. In design practice it extends past wording into locked references, seeds, weights and node graphs, so the same instruction produces the same output next month. **Locked style reference** — A fixed set of parameters, images and settings that every asset in a campaign is generated against, covering lens character, lighting direction, palette and how materials read. Without one, a set of forty images drifts and looks like it came from four different shoots. **LoRA** — Low-Rank Adaptation. A small trained add-on that teaches an existing image model a specific subject or style without retraining the whole model. In commercial design it is how a generated scene contains your actual product rather than something resembling it. **ControlNet** — A conditioning method that constrains where things appear in a generated image, using an input such as a depth map, pose or edge outline. It converts composition from something you hope for into something you specify, which matters when an image must fit a fixed layout. **Denoising** — The process by which a diffusion model turns random noise into a coherent image over a series of steps. Fewer steps is faster and rougher; more steps is slower and more resolved. It is the reason generated imagery emerges rather than being drawn. **Provenance** — What a model was trained on and whether those rights permit commercial use of its output. It is the deciding factor in regulated sectors, and it is why tools trained on licensed material get chosen over tools with better aesthetic range. **Design system** — A maintained set of components, tokens, spacing rules and naming conventions that a product's interfaces are assembled from. Its value is consistency as a product grows, which is why incoming designers should inherit one rather than propose a redesign. **Design token** — A named value, such as a colour or spacing unit, referenced by name rather than hard-coded. Changing the token changes every place it is used, which is what makes a rebrand a small job instead of a large one. **Upscaling** — Increasing an image's resolution after generation, usually with a model that adds plausible detail rather than simply interpolating pixels. Useful for print, and a common source of invented detail that needs checking before anything ships. **Editable source file** — The layered, working version of a deliverable, in a format such as Figma, Illustrator or After Effects, rather than a flattened export. Receiving sources is what lets your team change the work later without commissioning it again. **White label** — Work produced by one company and delivered under another's brand. In design placement it means an agency buys execution capacity, delivers it in their own files to their own client, and the supplier never appears. **Placement fee** — A charge for introducing a worker to an employer, typically levied when a contractor converts to permanent. Its size and timing determine how much freedom a client actually has to hire someone they like. **MLR review** — Medical, Legal and Regulatory review. The approval process pharmaceutical and healthcare material passes through before publication, requiring every claim to be traceable and every version accounted for. Design that ignores it gets rebuilt. ## Questions ### Why do these definitions differ from the tool vendors' versions? Vendors define a term by what their product does. We define it by what it changes about cost, risk and what you end up owning, because that is the part that affects a hiring decision. ### A designer used one of these terms differently. Who is right? Possibly both. The vocabulary in this field is unsettled and two people can say provenance and mean quite different things. Treat a mismatch as a question worth asking rather than a red flag. --- Hire AI Designer · https://hire-ai-designer.com/ · Figures on this site are the company's own and are not independently audited.