Furniture is the category where AI is supposed to fall apart. The reasoning sounds right: a chair has to sit in a room at a believable size relative to the sofa, the rug, the window, on a consistent perspective, with a real shadow where it meets the floor. Scale and perspective are exactly the kind of spatial reasoning that research has flagged as a genuine weak spot for text-to-image models, which offer limited control over the geometric structure of what they generate. So I expected furniture to be the test that broke on scale.
It did not. I ran one brief, a mid-century walnut lounge chair with tan leather upholstery in a sunlit living room, through four of the strongest image models with the same prompt: Nano Banana 2, GPT Image 2, Seedream 4.5, and FLUX.2 Pro. The scale held up on all four. The real catch was different, and it is the same lesson running through this whole series: each model produced a believable chair, but not the same chair, and not necessarily yours. This is the furniture entry alongside the skincare, jewelry, supplements, makeup, food and beverage, footwear, candles, and clothing tests and the broader best AI image model for product photography roundup.
Quick answer
- Best wide room scene with correct scale: Nano Banana 2 and FLUX.2 Pro. Both placed the chair believably in a full room with real proportions and shadows.
- Best cinematic hero and material: Seedream 4.5 (and the cheapest photoreal option) and GPT Image 2. The richest leather and walnut grain, in a tighter crop.
- The real limit: from a text prompt, all four designed a different mid-century chair, and FLUX.2 Pro redesigned the silhouette. The scene is yours to direct; the product is generic unless you feed a reference.
If you only remember one thing: AI furniture photography solves the room, not the SKU. For a mood or a concept, a text prompt is great. For your actual catalog piece, start from a photo of the real thing.
The test, model by model
One brief, four models, same prompt. I judged scale and perspective first (the feared hard case), then material, then how faithfully each kept the chair's design.
Nano Banana 2 produced the most usable room scene. The chair reads at the right size against the sofa, the sideboard, and the plant, the perspective is consistent, and the contact shadow on the rug grounds it. This is the scale test passing cleanly: nothing floats, nothing is comically large or small. For a catalog room shot, this is where I would start.
Seedream 4.5 made the most beautiful image again, exactly as it did on skincare, jewelry, and the others. The raking window light, the creasing in the leather, the walnut grain, it looks like a paid editorial shot, at the lowest cost of the photoreal models. The tradeoff is framing: it gave a tight hero crop rather than a wide room, so it is the better PDP lead image than catalog room scene. The chair geometry is coherent and the materials are the richest of the four.
GPT Image 2 matched Seedream on material and sat between the hero and the room in framing. The exposed walnut frame and the tan leather are rendered with real detail, and the chair's scale against the sofa behind it is believable. It is the most expensive model in the test, and here that buys material fidelity rather than a fundamentally different result on scale.
FLUX.2 Pro gave the cleanest wide room scene, believable scale, consistent perspective, a real shadow on the rug, and excellent walnut-shell grain, at the lowest cost. But it is also the clearest illustration of the real catch: it redesigned the chair into a curved shell-back wingback rather than the frame-and-loose-cushion mid-century chair the others rendered. The scale is right and the room is lovely, but it is a different chair. As elsewhere in this series, FLUX makes a strong overall image while taking liberties with the exact design.
The comparison
| Model | Scale and perspective | Material | Framing | Kept the exact design | Rough cost/image |
|---|---|---|---|---|---|
| Nano Banana 2 | Best, believable full room | Strong | Wide room scene | Faithful to the brief | ~9.3 credits |
| Seedream 4.5 | Good (tight crop) | Best, richest detail | Cinematic hero | Coherent | ~4.8 credits |
| GPT Image 2 | Good, believable vs sofa | Best-tie, rich | Styled mid-crop | Coherent | ~26.4 credits |
| FLUX.2 Pro | Best wide room | Strong | Wide room scene | Redesigned the silhouette | ~3.6 credits |
Credit costs are first-hand from this test on Masonry; per-image rates move, so check current pricing before you budget.
Why furniture is really an identity problem, not a scale problem
I went in expecting to write about scale. The interesting result is that scale is mostly solved, and the real constraint is one the other posts in this series keep surfacing.
Scale held, which was the surprise. The research is clear that proportion and perspective are a documented weak spot for text-to-image models, and for a complex multi-object scene that still shows. But for a single hero piece in a room, all four models placed the chair at a believable size, on a consistent perspective, with a real floor contact shadow. If you are staging one product in one room, the scale fear is largely outdated.
Identity is the real limit, because a prompt is a category, not a SKU. Asked for "a mid-century walnut lounge chair," each model designed its own plausible mid-century walnut chair, and they are all different chairs. FLUX went furthest and changed the silhouette entirely. For most of the products in this series, the identity risk is in a detail: a label, a shade, a logo, a print. For furniture the entire product is the design, the exact frame, joinery, cushion, and proportion, so a from-scratch generation does not give you a slightly-wrong version of your chair. It gives you a different chair that happens to share a style. That is fine for a concept board and useless for a catalog.
Which is why furniture, more than any category here, wants a reference image. The fix is the same one the clothing test pointed to: do not generate the product, generate the scene around the real product. Start from a clean photo of your actual chair and have the model restyle the room, the light, and the season while keeping the piece itself. The scale will hold, the room will look great, and crucially it will be your SKU in the shot.
How to shoot your furniture catalog without a room set
The workflow is the roundup approach, tuned for a product whose identity is its design. For concepts and mood, a text prompt across these models is fast and the scale is trustworthy. For your real catalog, feed a reference photo of the actual piece and restyle the room around it, so you get true product identity plus AI room variety. Judge scale and the floor shadow, which were reliable here but still worth a glance, and remember that vendor conversion and cost-saving percentages for room scenes are marketing numbers: room scenes do tend to outperform plain white backgrounds, but measure your own.
With the Masonry CLI you can do both from one place, generate concept rooms across models, or pass your real product as a reference to keep the actual piece while restyling the scene:
masonry image "mid-century walnut lounge chair in a sunlit minimalist living room, realistic scale, photoreal" --model gemini-3.1-flash-image-preview masonry image "place this exact chair in a warm sunlit living room, realistic scale and shadow" --ref ./real-chair.png --model seedream-4-5
The bottom line
Furniture turned out not to be a scale story. Every model placed the chair believably in the room, and the scale fear is mostly behind us for single-product scenes. The real lesson is the one this whole series keeps landing on: a text prompt gives you a category, not your product, and for furniture the product is the design. So use AI to own the room, and use a reference photo to keep the chair. Nano Banana 2 and FLUX.2 Pro for the widest believable rooms, Seedream 4.5 for the best hero at the lowest cost, all of them only as faithful to your SKU as the reference you give them. Run your own piece from one place with the Masonry CLI, or see how the same fidelity-first logic plays out across every product type in our best AI image model for product photography roundup.
