The difference between practices getting useful work out of generative tools and practices getting decorative noise is not the model they use. It is whether they have a method. Treated as a slot machine — type a description, generate, discard, repeat — the tool produces images that are individually attractive and collectively useless. Treated as an assistant that has to be briefed, corrected and checked, it becomes a legitimate part of concept work.
Three things separate the two. The model is given real geometry to work within, the design parameters are stated explicitly rather than implied by adjectives, and every output passes a fixed set of checks performed by a person. None of this is exotic. It is the same discipline applied to any junior contributor: brief properly, review consistently, and never let unchecked work leave the office.
Brief it the way you would brief a draftsperson
An instruction like warm, inviting restaurant lighting contains no information a designer would act on. The equivalent instruction to a person would be returned with questions, and if clarification is not requested or provided, the model is likely to invent the missing details.
The fix is to state the parameters that would appear in a specification.
| Parameter | Vague | Stated |
|---|---|---|
| Colour temperature | "warm" | 2700K in the dining area, 3000K in the bar working area |
| Distribution | "well lit" | Downlights over tables, wall grazing to the stone wall, no general ceiling wash |
| Beam behaviour | "spotlights" | Narrow 26° on each table, pool contained to the table surface |
| Contrast | "moody" | Tables brighter than circulation; ceiling left dark |
| Source visibility | — | No visible bright sources in a seated field of view |
| Mounting | — | Track at 2.4 m, pendants at 2m above table |
The second column describes the desired appearance. The third is what the space is, and it produces markedly better output because it removes the invented answers. It also has a secondary benefit: writing it forces the design decision to be made before the image is generated, rather than discovered afterwards by picking whichever render looked best.
One variable per iteration
The instinct when an output is disappointing is to rewrite the whole brief. It is the wrong move, because the next output differs in a dozen ways and there is no way to know which change helped.
Change one thing. Colour temperature, or beam angle, or the contrast relationship — one, then look. This is slower for the first three iterations and considerably faster after that, because it builds an understanding of how the tool responds that carries to the next project.
Keep the versions. A scheme that was heading in a good direction four iterations ago is recoverable if the intermediate outputs still exist, and gone if they were discarded. Name them by what changed rather than by number.
Annotate rather than re-describe
Once an output is close, prose stops being the efficient channel. An arrow drawn to where a beam should land, a circle around a fitting that should not be there, a note on the surface that should stay dark — marking up the image communicates a spatial correction spatially, and it survives the round trip intact. The same correction written as a sentence tends to be applied somewhere else in the frame, or applied everywhere at once.
This is also how the correction gets recorded. An annotated image is a reviewable artefact; a paragraph in a chat window is not.
The checks a person performs, every time
The review is fixed and it happens before anything is shown to anyone.
Against the design. Do the fittings in the image correspond to the schedule? Is the light patch the size that beam angle produces from that mounting height? Do the shadows agree on where the sources are? Is a source shown in a position where an occupant would look directly into it?
Against verified performance requirements. The image is reviewed to ensure that it does not imply verified illuminance, uniformity, glare performance or lighting power density. These outcomes must be established through appropriate lighting calculations, verified project data and assessment against the applicable Australian Standards and NCC requirements. If a client's question can only be answered with a number, the generated image is not the appropriate place to answer it.
Against reality. Is every fitting depicted something that can be bought, mounted where it is shown, and connected? Generated imagery is fluent at producing luminaires that do not exist and mounting conditions that no ceiling would accept.
The line
The boundary is not a matter of taste, and it is worth writing down so that it is the same on every project.
Concept imagery, mood studies, option exploration, client conversation — AI belongs here, and the method above is what makes it reliable. Anything that becomes a number someone relies on belongs to calculation. For example, in an interior lighting project: illuminance and uniformity to AS/NZS 1680, cable sizing and protection to AS/NZS 3008 and AS/NZS 3000, energy compliance to Section J7 of the NCC, emergency provisions to AS/NZS 2293. Those are computed, checked and signed.
An AI-produced output may be carried forward when it provides a credible starting point, aligns with the design intent and can be professionally verified; it is discarded entirely when fundamental spatial, lighting or technical errors make correction less reliable or efficient than starting again.
What this adds up to
The practices that will get the most out of these tools over the next few years are unlikely to be the ones with the newest model. They will be the ones that treat generated output as a draft requiring review, keep the calculation exactly where it has always been, and are honest with clients about which of the two they are looking at.
That is a low-drama conclusion, and it is the one the work supports. The tool compresses the front end of a project meaningfully. It does not touch the part where the design has to be right.
FAQ
How do you get consistent results from AI across a project?
State the design parameters explicitly — colour temperature, beam angle, mounting height, contrast relationships — rather than describing the mood, work from the same reference geometry for every view, and change one variable at a time so the effect of each change is legible.
Should a lighting practice use AI at all?
For concept visualisation and option exploration it is straightforwardly useful and already common. For final calculation, specification and compliance documentation, generative AI should not be relied upon, and presenting generated imagery as if it demonstrated performance is the one thing to avoid.
How do you stop AI imagery from setting unrealistic client expectations?
Check every image against the actual design before it is shown — fitting positions against the layout, light patch size against the specified beam angle and mounting height — and label concept imagery as concept imagery. An image that depicts an effect the specified fittings cannot produce creates a problem that surfaces at handover.
Does this method work for anything other than lighting?
The structure does: supply real geometry, state parameters rather than adjectives, iterate one variable at a time, and fix the checks a person performs. The parameters are what change between disciplines.
SL Design Consultant is an independent lighting and electrical design consultancy in Sydney, delivering lighting and electrical design across commercial, hospitality, retail, heritage and public projects. Our product-neutral recommendations are guided by design intent, technical performance and project requirements.