AI Room Visualisation From Client Photos: A Designer's Workflow

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for AI Room Visualisation From Client Photos: A Designer's Workflow.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for AI Room Visualisation From Client Photos: A Designer's Workflow.

Upload two or three level, well-lit photos of the client's room to a tool that edits images, such as ChatGPT, Gemini or Interior AI. List what must stay fixed and what should change, then refine in small rounds. Expect 20 to 45 minutes per room, and treat the result as direction, never as a measured drawing.

The catch is that these tools don't edit a photo the way a retoucher does. They redraw the whole scene from your picture plus your words, so a window can creep 20 centimetres sideways, a ceiling can gain height and a sofa can shrink to fit a wall it would never fit in real life. The workflow below is built around that weakness: you decide what is locked before you prompt, you check every output against the original, and every image leaves the studio captioned as a concept.

Follow me on Instagram@sagnikteaches

What the model keeps from a client's photo, and what it changes

Knowing where the drift happens tells you where to look when you review. The pattern is predictable enough to plan around:

Connect on LinkedInSagnik Bhattacharya
ElementWhat tends to happenHow to catch it
Viewpoint and room shapeUsually holds if you ask for "the same viewpoint"Compare the corner lines of floor and ceiling
Windows and doorsPosition roughly kept; size, glazing bars and door swing changeOverlay the original at 50% opacity
Ceiling height, beams, cornicingHeight grows, mouldings get simplified or vanishCheck the gap between door head and ceiling
Radiators, sockets, switches, ventsDisappear, because a clean wall looks betterTick each one off against the photo
Furniture scaleInvented; pieces are sized to look good, not to fitCheck the key pieces against your tape measurements
ColourShifts with the model's lighting; screens differ from paint cardsAlways follow with physical samples on site

Hold on to that last column. Everything later in the workflow exists to make those checks quick.

Subscribe on YouTube@codingliquids

Stage 1: Shoot photos with the model in mind (about 10 minutes)

Most bad AI visuals start as bad source photos. The model copies distortion, clutter and colour casts faithfully, then builds your new scheme on top of them. On the initial visit, take the photos for the AI as a separate job from your record shots:

  • Use the standard lens, not the ultra-wide. The 0.5x setting on a phone bends walls outwards, and the model treats the curve as real architecture. Stand in a corner at chest height instead.
  • Hold the phone level. Turn on the camera grid and keep vertical lines vertical. Tilted verticals come back as leaning walls.
  • Shoot three angles per room: the main view the client will judge the scheme from, the opposite corner, and a straight-on shot of the feature wall.
  • Use daylight with the lights off. Mixed warm bulbs and cool daylight give the model a colour cast it will "correct" in unpredictable ways.
  • Clear what you can. A pile of coats on a chair becomes a strange sculpture. Ask the client to clear surfaces before the visit.
  • Measure the things the model gets wrong: wall lengths, ceiling height, sill and window-head heights, and the distance between window and door on each wall. One photo with an extended tape measure against the wall gives you a scale reference for later.

Some rooms have no useful daylight: a basement snug, an internal shower room, a north-facing box room in winter. Don't mix light sources to brighten them. Switch on every light in the room, make sure the bulbs share a colour temperature (swap or unscrew the odd one out for the shot), and say so in the prompt: "lit only by warm ceiling spotlights; no window light". Otherwise the model invents a window's worth of daylight, and the client sees a bright, airy room that will never look like that at four o'clock on a winter afternoon.

Before any photo goes into a tool, crop or blur family photos, post, screens and anything that identifies the address, and switch off location data when you export. Ask for permission in writing. Something as short as this works:

We'd like to use photos of your [room] in an AI image tool to produce early
concept visuals for your project. We'll remove personal items from the photos
first, use them only for your project, and use a tool set not to train on
our uploads. Are you happy for us to do that?

Only send the last sentence if it's true for the tool and plan you use. Check the model-training switch in your account's privacy settings; business plans of the main assistants don't train on your content by default, while consumer plans usually need the switch turned off. If you work with client files regularly, the routine in anonymising client data before it goes into AI applies to photos too.

Stage 2: Choose the tool by the job in front of you

There are three kinds of tool worth knowing, and most studios end up using two of them:

Tool typeExamplesBest forWatch for
General assistant that edits photosChatGPT, Gemini (its image model is Nano Banana 2)Fast restyles, "what if" material swaps, iterating in conversation, combining the room with product photosUsage limits vary by plan and change often; Gemini marks images with a visible watermark and an invisible SynthID mark
Dedicated room-redesign toolInterior AIMany style variations of one photo, structure-keeping modes, staging empty roomsResults can look generic; as of September 2026 its Pro plan lists at $29 a month billed yearly or $49 monthly for 1,000 designs
Photo editor with generative fillPhotoshop's Generative FillLocal changes: swap one chair, re-tile a splashback, remove clutter without redrawing the roomSlower, and it needs someone comfortable in an editor

The practical pairing for most studios is a general assistant for exploring directions and a photo editor for the final tidy-up, because local edits leave the rest of the photo untouched. Interior AI states that it doesn't train on uploaded photos, which makes the client-permission conversation simpler. If you already build SketchUp models, rendering from the model beats editing a photo for accuracy; the comparison in AI renders versus outsourced visualisation covers when that switch pays off.

On cost, the subscription is rarely the expensive part. At Interior AI's monthly Pro price, $49 for 1,000 designs is about 5 cents a design, and a studio generating 60 variations a month is paying for capacity it never touches. The real cost is designer time: at 20 to 45 minutes a room, a three-room project takes one to two hours of prompting and checking, which is why the lock list and the overlay check matter more than which tool you pay for.

Stage 3: Write the prompt as a brief with a lock list

"Make this living room Scandi" produces a pretty picture of a different room. The prompt that works reads like a brief to a junior designer: what stays, what changes, what style and light, and what to do when something won't fit. Copy this and fill in the brackets:

This is a photo of a client's [room type]. Create a concept visual of the
same room from exactly the same viewpoint.

KEEP EXACTLY (do not move, resize or remove):
- window positions and sizes, including the glazing bars
- the door on the [left] wall and its position
- ceiling height, [the beam / the cornice / the fireplace]
- the radiator under the window, sockets and switches
- the [oak floorboards] and the direction they run

CHANGE:
- walls: [warm off-white], matt finish
- replace the sofa with a three-seat sofa in [rust velvet], low arms,
  about [210] cm wide
- add a round [jute] rug about [200] cm across under the coffee table
- add a [brass floor lamp] to the right of the sofa

STYLE AND LIGHT:
- [relaxed mid-century], lived-in rather than showroom
- soft daylight from the window, matching the original photo
- photographic, not illustrated; no people; no text or logos

Add nothing that isn't listed. If an item won't fit at the size given,
leave it out rather than shrinking it.

Three details make the difference. Real dimensions give the model something to scale against. Naming the services (radiator, sockets) stops them disappearing. The final line matters most, because the model's default is to make everything fit by quietly shrinking it. If you have the actual products, attach their photos alongside the room photo and say "use the sofa in image 2"; both ChatGPT and Gemini accept several images in one message.

Commercial rooms need extra lines on the lock list. Picture a delicatessen that wants a six-seat café corner beside its chiller counter. The lock list there includes the counter, the extraction hood, the fire exit sign above the back door and a clear route from door to till. In an illustrative first attempt at a room like that, the model tidies the exit sign away because it spoils the look, and it narrows the route between tables to squeeze in a banquette. Neither would survive a fit-out, so both go on the lock list by name.

Stage 4: Refine one change per round, then stop

Treat each reply as a draft to mark up. The first round sets the overall scheme. In the second, fix drift in plain words. The third round is for finishing details. Asking for three fixes at once usually gets you two fixes and a new problem.

Here is what a typical first round looks like when you mark it up. Using the living-room brief above, an illustrative first output might come back like this:

  • Rust velvet sofa and round jute rug: correct, and roughly the right size.
  • Brass floor lamp: present, but on the left of the sofa, not the right.
  • Radiator under the window: gone.
  • Window: about a hand's width taller than in the photo.
  • A large rattan pendant: added, although nobody asked for one.

Pick the worst problem first, which here is the window, because it changes the architecture. The round-two prompt is one instruction plus a guard:

The window is taller than in the original photo. Make it exactly the size
and position it is in the original. Change nothing else in the image.

Round three restores the radiator the same way, and round four removes the pendant and moves the lamp. Four short rounds beat one long correction every time.

If three rounds haven't cured a drift, start a fresh conversation with the original photo and a tightened lock list. Long edit chains tend to soften detail and wander further from the source with each regeneration. For a single stubborn object, stop asking the assistant and make a local edit instead: select the chair in a photo editor and describe the replacement, so the walls and windows stay pixel-for-pixel.

The fastest review trick is the overlay. Put the original photo and the AI image on two layers in any editor that supports opacity (Canva, Photoshop and Affinity all do), set the top layer to 50%, and flick it on and off. A window that moved, a radiator that vanished or a ceiling that rose shows up in seconds, far faster than comparing two images side by side.

Stage 5: Caption and present it as a concept

A client who sees a photoreal image of their own room reads it as a promise. The caption is what keeps it a concept. Put this, or your version of it, on every image you share:

Concept visual created with AI from a photo of your room. For mood and
direction only: proportions, colours and products are indicative and not
to scale. Final selections will be confirmed in your specification.

The caption isn't decoration. A common way this goes wrong: a client forwards an uncaptioned concept to their joiner, who prices a run of shelving from the picture. The shelves in the image are a model's guess, 30 centimetres too long for the alcove, and the quote comes back high. The designer then spends a meeting explaining that the picture was never a drawing. With the caption on the image file itself, not just in the covering email, that conversation happens before the quote rather than after it.

The other question clients ask is why their room looks bigger in the picture. Usually the model has raised the ceiling a little or pushed the far wall back so the sofa fits with room to spare. Answer it plainly: "The image makes the room look larger than it is. The real room will feel cosier, which is why we've chosen a lower sofa and a smaller rug." Then run the overlay again, find the drift and correct it before the next meeting, so the version the client keeps on their phone is honest about the space.

Present in a fixed order: the original photo, the concept, then the real samples and products that would deliver it. Two or three directions are plenty; twenty variations turn a design decision into scrolling. The presentation side, including how concepts sit with sample boards, is covered in using AI for mood boards and presentations. Name versions clearly (room, direction, round) so that "the second one from Tuesday" can be found when the client asks for it three weeks later.

A wine merchant's tasting room, from photo to sign-off

To put numbers on the stages, take an illustrative job: a two-person studio refreshing the back room a wine merchant uses for evening tastings. The room is about 5.2 by 4.1 metres, with one sash window, a door on the opposite wall and an exposed brick wall the owner wants to keep.

  1. Site photos and measurements, 15 minutes. Three angles in daylight, tape photo against the brick wall, sill and door heights noted.
  2. Three directions in a general assistant, 35 minutes. A warm cellar look, a pale gallery look and a bar-style layout, each with the brick wall, window and radiator on the lock list. Two rounds of drift fixes, mostly the window growing taller.
  3. Local fixes in a photo editor, 20 minutes. Restoring a light switch and swapping an invented pendant for the real fitting the studio intends to specify.
  4. Captions and a short deck, 15 minutes. Original, concept, product board, for each direction.

That is about an hour and a half for three directions, on a subscription the studio already uses for other work. The overlay check earned its keep: direction three showed a floor-to-ceiling bottle wall about 2.4 metres wide, but the tape notes put the wall between window and door at 1.9 metres. Caught before the meeting, it became a talking point about a narrower rack rather than a disappointment at the specification stage.

Five ways client-photo visuals mislead, and the check for each

  • Mirrors and glass. Reflections show a room that doesn't exist, sometimes with an extra window. Remove mirrors from the prompt or cover them with a local edit.
  • Vanishing services. A wall without its radiator, sockets or thermostat looks cleaner than it will ever be. Tick services off against the original on every image.
  • Scale that flatters. Pendants come out too small, rugs too big and sofas too slim. Check each hero piece against your measurements before the client sees it.
  • Colour that only works on screen. Paint and fabric colours shift with the model's lighting. Never let a client choose a colour from the visual; use it to choose a direction, then test samples on the wall.
  • Products that can't be bought. Clients ask where to buy the lamp, and the model made it up. Put real products on the board next to every concept so the question has an answer.

Rights deserve a moment too. If the client's walls hold artwork, avoid restyling it into something new that you then publish in your portfolio, and ask before using any client-home visual in marketing. The wider picture on ownership and permissions is in faster concepts without rights issues. For estate-agency style staging of empty rooms, which carries its own disclosure expectations, see AI virtual staging for property listings.

Questions designers ask before using client photos

Is it safe to upload photos of a client's home to an AI image tool?

Only with the client's permission and after removing personal items such as family photos, post, screens and anything showing the address. Check the tool's training setting: business plans of the main assistants don't train on your content by default, consumer plans usually have a switch you must turn off, and some dedicated room tools state they never train on uploads. Keep the originals in your project folder, not in a personal chat history.

Can I charge for AI concept visuals?

Yes, as design time. The value is your brief, your product knowledge and the checking, not the button press. Many studios fold two or three concept directions into the concept stage fee rather than pricing images individually, which avoids clients asking for twenty free variations. Be open about how the visuals are made so the fee never looks like it was hiding a shortcut.

What if the client falls in love with something in the image that doesn't exist?

It happens, usually with a lamp, a rug or a joinery detail the model invented. Say so straight away and show the nearest real product next to it. Presenting a real product board alongside every concept stops the problem, because the client learns early that the image sets direction and the board sets what gets ordered.

Further reads

Sources: Interior AI product and pricing page; Google Gemini image generation overview (Nano Banana 2, SynthID and visible watermark); OpenAI help article on images in ChatGPT (editing uploaded photos).

Want AI visuals built into your concept stage properly?

On a 1:1 call we'll look at how your studio presents concepts now, pick the tool that fits your projects, and set up a prompt, checking and captioning routine your team can repeat.

Book a 1:1 call with me