Paint swatches lie. A colour that looks warm and grounded on a 50mm card goes nuclear across four walls, and by then you have spent a weekend and £200 finding out. The same goes for a sofa you fall for in a showroom that turns out to eat half your floor. Seeing the room finished — actually finished, in your own light, with your own windows and your own awkward radiator — is the thing that stops bad decisions. That is the job AI interior design does well, and it does it in about ninety seconds per version.
What follows is the workflow, in the order you will hit each decision: how to shoot the room, how to decide what you are actually changing, how to write a prompt that produces furniture you could buy, how to iterate without losing the room, and how to turn the result into something a contractor or a shopping list can use.
Shoot the Room Properly Before You Touch Any Tool
The single biggest predictor of output quality is the photo you feed in. A dim, tilted phone snap taken at chest height produces a redesign that looks like a different house. A clean, well-lit shot produces something you can hold a tape measure against.
Stand in a corner and shoot towards the opposite corner. That diagonal gives the model two walls, the floor plane and usually the ceiling line, which is what it needs to understand the volume of the space. Shooting flat at a single wall gives it almost nothing to work with and the results drift.
- Shoot in daylight, mid-morning or mid-afternoon. Overhead bulbs cast an orange tint the model will faithfully reproduce in every version you generate.
- Hold the phone at roughly 1.5 metres, chest to eye height, and keep it level. Tilting up or down warps the vertical lines and the model compounds the distortion.
- Clear the obvious clutter — laundry, cables, the drying rack. You do not need to empty the room, but anything genuinely temporary should go.
- Take three shots from three corners. You will want to test the same scheme from more than one angle later, and matching schemes across angles is far easier when the source photos are consistent.
- Turn off portrait mode. Artificial background blur confuses depth estimation and you lose detail on the far wall.
If the only usable photo you have is old, low-resolution or shot in bad light, run it through an image upscaler first. A 1024px-wide source produces mushy furniture edges; a 2048px source produces chair legs you can count.
Deal With Clutter You Cannot Move
Sometimes the thing in the way is a stack of boxes you are not shifting for a photo, or a piece of furniture that is definitely going but is still there. Strip it out before you redesign, not during. An object remover handles this cleanly, and the model then fills the empty space on its own terms rather than trying to design around a wardrobe you have already sold. There is a fuller walkthrough of that process in this guide to removing objects without reshoots.
Decide What Is Fixed and What Is Up for Grabs
This is the decision most people skip, and it is why their results feel useless. A render that moves the window, widens the room and installs a skylight is a mood board, not a plan. Before you write anything, split the room into three lists.
Fixed: window positions, door swings, radiators, ceiling height, structural walls, the fireplace. Name these in the prompt explicitly so the model preserves them.
Negotiable: flooring, wall colour, built-in storage, lighting fixtures. These cost real money but they are genuinely changeable.
Free: furniture, rugs, textiles, art, plants. This is where most of a room's character lives and where you should spend your iteration budget.
If your actual question is "what colour should these walls be", do not generate a whole new room. Lock everything else and change one variable. You will get a straight answer instead of five attractive but incomparable images.
Write the Prompt Like a Brief, Not a Wish
Generic prompts produce generic rooms. "Modern living room" gets you the same grey sectional and fiddle-leaf fig everyone else gets. The fix is specificity across four axes: style, palette, materials, and what stays.
A weak prompt:
Make this living room look modern and cosy.
A prompt that produces something you can act on:
Redesign this living room in warm mid-century modern. Walnut furniture, a low three-seat sofa in oatmeal bouclé, a muted terracotta and sage palette, brass floor lamp, wool rug with a subtle geometric pattern. Keep the existing window position, radiator and door frame exactly as they are. Natural afternoon light from the left. Photorealistic, architectural photography.
Five things changed there: a named style, named materials, a defined palette, explicit preservation instructions, and a lighting direction. Each one removes a decision the model would otherwise make for you — usually badly.
Name Materials, Not Adjectives
"Cosy" is a feeling. "Bouclé, wool, unlacquered brass, oiled oak" is a specification. Models have seen millions of captioned interiors and they respond far more reliably to material nouns than to mood adjectives. Build a short vocabulary you reuse:
- Textiles: bouclé, linen, velvet, jute, wool flatweave
- Woods: walnut, white oak, oiled ash, reclaimed pine
- Metals: brushed brass, blackened steel, polished chrome, antique nickel
- Stone and surface: honed marble, microcement, terrazzo, limewash plaster
Two or three materials per prompt is the sweet spot. Six turns into visual noise and the model starts inventing hybrids that nothing on the market resembles.
Give the Light a Direction
Interior renders fail most visibly on lighting. If your real room gets north light that stays cool all day, a prompt asking for "golden hour sun streaming in" produces a room you will never actually see. State the direction and quality you have: soft diffused north light from the right, overcast day or low warm sunlight through west-facing window, long shadows. The scheme you choose should look good in your light, not in borrowed light.
Iterate on One Variable at a Time
Once you have a base render you like, resist the urge to re-prompt from scratch. Conversational editing tools let you change a single element and keep everything else intact, which is what makes comparison possible.
With a natural language image editor, you can take a finished render and say:
Change the rug to a flat sisal weave and keep everything else identical.
Then:
Replace the brass pendant with a white paper globe pendant.
Now you have three images that differ by one decision each. Put them side by side and the right answer is usually obvious within seconds. That is impossible when every generation reinvents the whole room.
For wall colour specifically, this is the fastest real-world test available. Generate the same room in four shades, view them on a phone in the actual room at the actual time of day you use it, and you will catch the one that goes green in the evening before you buy five litres of it.
Keep a Version Log
Save the prompt text alongside each image. Three days in you will have forty renders and no memory of which one used the oak flooring versus the ash. A plain text file with v07 — oatmeal sofa, sisal rug, sage walls is enough. It also means you can rebuild a near-miss instead of hunting for it.
Test the Scheme Against the Rest of the House
A room that looks perfect in isolation can clash badly with the hallway it opens onto. Once you have a scheme, generate the adjacent spaces using the same material vocabulary and palette. Reuse the exact phrasing — same wood, same metal, same two or three accent colours — so the renders read as one house rather than a showroom tour.
This is also where you check the scheme at different times of day. Run your favourite render twice more, once as early morning, cool light and once as evening, lamps on, no daylight. Plenty of schemes that sparkle at noon go flat and grey at 7pm, and a room you only use in the evening needs to be judged in evening light.
Turn the Render Into a Shopping List
A render is not a plan until it maps to things you can buy. Work through it piece by piece and classify each item.
- Findable: the model has produced something close to a real, commercially available product category. A low walnut-framed three-seater exists; search on the material and silhouette, not the render.
- Approximate: the shape is useful but the specific piece is invented. Use it as a brief for what to look for — "round side table, 45cm, dark stained wood".
- Impossible: the model has built a bookshelf that floats or a sofa with five legs. Ignore it. Models still fudge joinery, reflections and the geometry of anything with repeated parts.
Check the scale before you commit. Models are poor at absolute dimensions and will happily place a 2.4m sofa in a 2.1m alcove. Take the render, note the real width of one fixed element — the window, the chimney breast — and use it to sanity-check everything else against it. If the sofa reads as 1.3 window-widths and your window is 1.6m, you are shopping for a 2.1m sofa.
For anything you are presenting to a contractor, builder or landlord, push the resolution up. A higher-resolution export prints legibly at A3 and survives being zoomed into on a tablet on site. A 1024px render does neither.
Know Where the Tool Stops
AI interior design is excellent at style direction, palette testing and showing you what a room could feel like. It is unreliable at anything measurable. It does not know your ceiling height, your joist positions, where your soil pipe runs, or whether that wall is load-bearing. Treat every render as a visual brief, never as a drawing.
It also tends towards a house style — warm neutrals, a lot of bouclé, plants in the corner — because that is what dominates its training data. If you want something genuinely unusual, you have to push much harder in the prompt and expect more rejected attempts. Naming a specific period, a specific designer's material language, or an unusual palette combination gets you further than asking for "unique".
Start with one room, one angle, and three schemes. Pick the one you keep coming back to after a day away from the screen, then change exactly one thing about it and see whether it gets better. That single discipline — one variable, one comparison — is what turns a pile of pretty renders into a decision you can actually spend money on. If you want to build out the rest of your workflow, the full set of tools covers everything from cleaning up the source photo to producing presentation-ready exports.
