A furniture brand sent us a forty-second film it had generated in-house. In the first shot a three-seat sofa sat under a tall window. In the second shot the same sofa had a narrower armrest, the window had slid to the left, and the floorboards ran the other way. Nobody in the company called it ugly. They called it wrong, and the sales team noticed before a single customer did.
Furniture and interior work is judged on exactly the things generative video is weakest at: proportion, scale, and the behaviour of materials under light. A chair is not a mood, it is a set of dimensions. This is the practical route we use at KAMOD when a furniture or interior brand wants AI film that a customer will trust.
Why rooms fall apart between shots
A video model does not hold a 3D model of your sofa. Each generation invents the room again from the prompt and the reference frame, which is why a sofa can gain an arm, a rug can change size, and a pendant light can move half a metre between cuts. Fast camera moves make this worse because they expose the parts of the room the model had to invent.
The fix is editorial, not technical. Cut less. Hold the camera. Reuse one environment across the whole film and treat every new angle as a risk you have to earn. In a typical 30-second furniture piece we use four to six shots total, and two of those are close detail shots where the room behind the product is out of focus and cannot contradict itself.
Photograph the product, generate the world
The single biggest quality decision is what you let the model invent. Our rule: the product is photographed or supplied as a real asset, and only the environment around it is generated. A sofa photographed on a clean sweep in the studio keeps its exact arm shape, seam line and leg profile. Drop that cut-out into a generated room and the customer sees your actual product in a space they can imagine living in.
The reverse — generating the product too — works only for concept films, mood reels and internal pitches, where nobody is going to measure the armrest against a spec sheet. Say which of the two you are buying before the job starts, because the price difference is real and the credibility difference is larger.
Build one room and stay in it
For a collection film, we build a single environment plate first: one room, one time of day, one light direction. We approve that plate as a still with the client before any motion is generated. Every subsequent shot is generated from that plate plus a fixed camera description, so the window stays where the window was.
Two practical details matter more than they sound. First, lock the light direction across the whole film — if the window light comes from the left in shot one, it comes from the left in shot six, and any shot that breaks it gets cut. Second, keep the floor visible in the wide shots; floorboards, tiles and rugs are the strongest continuity cue a viewer has, and they are the first thing that betrays a regenerated room.
Material truth: textile, veneer, brushed metal
Close-ups are where furniture brands either win or lose the viewer. A fabric weave that turns to mush, a walnut veneer whose grain repeats every ten centimetres, a brushed steel leg that reflects a room you cannot see — all of these read as fake instantly to anyone who has touched the real piece.
What helps: shoot the close-ups for real. A five-second macro of the stitch line, the joint, the drawer runner is cheap to film and impossible to generate convincingly. Save the generated footage for wide room shots, transitions and atmosphere, where the viewer is looking at the space rather than inspecting the joinery. When you must generate a material close-up, keep the camera slow, keep the light soft and directional, and avoid any shot where the surface has to reflect a specific object.
Scale needs a human reference
Furniture mistakes are usually scale mistakes. A dining table that seats six but looks like it seats four, a sideboard that comes up to chest height, a bed that fills a room it could never fit in. Generative shots drift on scale because nothing in the frame defines a known size.
Put a person, a door frame or a standard 2.4-metre wall in the wide shot. If a human appears, brief the pose precisely: standing beside the table with a hand on the backrest, sitting at the desk, reaching into the wardrobe. A model that has a human at a known height will place the furniture more plausibly, and the viewer gets the size cue without being told the dimensions.
What to shoot for real, what to generate, what to skip
The honest split for a collection launch of, say, twelve pieces: photograph all twelve products in the studio, film real macro details of the three hero pieces, generate four to six room environments, and cut a 30-second master plus vertical variants for social. That is roughly two studio days plus a week of post, and it produces a film that survives contact with a customer who already owns the sofa.
Skip the temptation to generate the whole catalogue as one continuous walkthrough. Long single-take moves through a generated house look impressive for four seconds and then demand continuity the model cannot hold. If you need a walkthrough, build it as separate rooms cut together with matched light.
Trade-offs worth stating plainly: generated environments cost less than renting and dressing a real apartment, but they cannot show the exact fabric in the exact daylight your customer will have at home, and they will never replace a real photograph of the actual product. The hybrid is not a compromise, it is the only version that is both affordable and true.
If you are planning a collection film, a showroom loop or a marketplace video set, send us the product list, the room references and the delivery specs. We will tell you which shots we would film and which we would generate before you commit a budget. KAMOD is an AI video studio in Tashkent, working with brands worldwide.