Making a generator keep the size

Question: how do you set the size

A furniture try-on lives or dies on one number. A 192 cm sofa has to come out 192 cm wide, otherwise someone buys a thing that will not fit. Image generators do not promise that: they draw an object at whatever size looks good in the frame.

We looked for a way to hand a model the size so that it actually keeps it. Below is the protocol and three experiments: what to set the size with, which models read it and whether the result repeats across rooms.

1. 01Latest: how to check the size
2. 02The instrument: segmentation, not opinion
3. 03A correct frame made the result worse
4. 04Question: how to set the size
5. 05Method: the door as a ruler
6. 06Experiment 1: three ways to say "110 cm"
7. 07Experiment 2: eleven models, one frame
8. 08Experiment 3: does it repeat elsewhere
9. 09Where the method breaks
10. 10Previous round: 14 products, 4 models

Question: how do you check the size

Earlier rounds looked for a way to give the model a size. This one asks how to verify the model kept it. Until September we had nothing but our own confidence: we showed the finished shot to a text model and asked where the object was.

In the 8 September run all twelve checks came back "unverified". The breakdown showed the coordinate format was not the problem.

A text model cannot see pixels

The reviewer was shown the finished shot and the target frame, and asked for the object bounds as fractions of the image. It matched the frame instead of measuring what was drawn.

The model does not measure, it invents something plausible. While the target frame is in front of it, it converges on that frame — that is, it confirms itself.

The instrument: segmentation, not opinion

The object is found by text-prompted segmentation, and the target frame is never shown to it — there is nothing to conform to. Height and centre are computed in code. The wording of the request turned out to decide everything.

A product code or a chain of adjectives kills the search as reliably as a typo. A short noun phrase works. The instrument was checked against two hand measurements and matched both exactly; every measurement now runs two phrasings, and if they disagree by more than 10% the check refuses to answer.

What the instrument found

The first thing the instrument showed was that the error was not in the generator but in our own arithmetic. Scale comes from the depth at a reference point, and that point was landing on a foreground object instead of the floor.

The "truth" column was obtained independently: in a room measured with a tape, scale was derived from an object of known size and transferred to the placement point by a depth ratio. Two anchors at different distances gave answers one pixel apart, and the depth ratio matched the physical ratio of scales to within 1.1%.

A correct frame made the result worse

Three new products, one room, the same tap. Exactly one thing differed: in the second arm the scale came from a measurement against an object of known size rather than from the model — that is, the frame was nearly correct.

The generator draws the object 1.10–1.32 times larger than the frame, and the factor does not depend on where the frame came from: it is a property of the product, not of the arithmetic. So today’s decent accuracy rests on two errors cancelling — an undersized frame and an oversized drawing. Remove one and the other shows in full.

What this round does not prove

• One room, three products, one shot per arm. A direction, not a statistic.
• The mask does not hold the size: on the path with a real alpha mask, 29% of the pixels outside the mask changed anyway — the model redraws the whole frame and treats the mask as a hint.
• The measurement only understands English product names and refuses to work when two phrasings disagree. We treat refusal as correct behaviour: better silence than a confident wrong number.
• The frames from this round were shot in a private home and are not published. The section runs on tables, as every round has since 31 August.

The door leaf as a ruler

To check a size in a photo you need the scale of the room: how many pixels fit into one centimetre. We take it from the interior door, whose leaf is 210 cm almost everywhere and usually visible in full. That measurement is 1–2% off. The older method, floor-to-ceiling distance, was off by 13 to 43%.

The product dimensions from the shop listing then become a rectangle in pixels, drawn straight onto the room photo in red: that is the correct size. The model gets the same photo with a green frame and a request to draw the product inside those lines and remove the frame. A hit is checked by overlaying the red reference on the result.

Three rooms shot on a phone, 1280 × 960. Products are real IKEA and Ozon listings with their own dimensions. A hit means the top, bottom and sides of the object sit inside the frame, visually within 10%. Each model ran each case once, with no cherry-picking. The rooms are a private home, so the raw frames stay unpublished; every measurement and outcome is in the tables below.

Three ways to say "110 cm"

One product, a plant 110 cm tall, one room, one model — Seedream 4.5. The only thing that changes is how the model learns the size: from the prompt text, from a cut-out pasted into the frame, or from a rectangle drawn on the photo.

The model does not trust words at all. A pasted cut-out sets the size geometrically, yet the model still redraws the object bigger. A rectangle inside the picture turned out to be the only language that works.

Eleven models, one frame

With the method settled, the question moves to the models. A bookcase 80 × 237 cm, the same room and frame, the same prompt word for word. We tested everything on Replicate that accepts a product photo alongside a room photo.

Two models out of eleven read the frame. The rest fail in different ways, most often in the same direction: the product comes out bigger than asked. Qwen is its own case, drawing the bookcase correctly but leaving the green frame in the picture.

Does it repeat in other rooms

A single good shot proves nothing. The same two finalists, plus Wan 2.7 as the cheapest option, ran four new cases in two other flats: a 192 × 94 cm sofa, a 270 cm storage unit deliberately wider than the wall, an 80 × 237 cm bookcase and a 110 cm plant. The room from experiment 2 is not repeated here.

The losers err in the same direction, which matters more than the score: this is not randomness but how the model reads the rectangle. It treats it as a hint about position, not a boundary. All three missed on the plant: a living crown has no clean edges, and the frame stops working.

Where the method breaks

A rug lies in the floor plane, so it gets a perspective footprint instead of a rectangle. The 250 × 250 cm rug landed on that footprint correctly, and then a rolled-up tube appeared beside it that nobody asked for. The lamp frame came out narrow, 109 pixels in a 1280 pixel frame, and the model simply missed it.

What is fair to say about these numbers: the rooms come from one flat, the products from two shops, the hit is judged by eye, and each model ran each case once. This is not statistics, it is a tool-selection protocol. We publish the method and every outcome, misses included, so the reasoning can be rechecked; the illustrated round on a neutral room is in the next section.

What the lab ran since

Short takes from the newest rounds — only the conclusions that survived scrutiny. Full protocols with every frame live in the research repo.