What the model knows about space — and what it does not
The image model turns your text into a condensed bundle of meaning and builds the picture from that. What survives that bundle very well is what is in the picture: product, table, window, light. What survives badly is the relation: what is behind what, what is to the left of whom, which one is higher up. The model saw on its training images that a cup stands on a table, but from the sentence "the cup to the left of the laptop" it cannot reliably work out which side that is.
This is the explanation behind an everyday experience. If you have ever tried to ask for an image where you know exactly where the logo goes and where the caption needs room, you have lived through it: the picture is handsome, the contents are right, only the arrangement is random. The good news is that the situation is not hopeless — it just cannot be solved with words.
The measurement that shows the limit
In January 2026 a benchmark built specifically for spatial arrangement was published: 1230 prompts, 21 models, examining nothing but whether the thing ends up where it was asked for. The result is sobering.
- Even the best models score around 60 percent overall — meaning every third or fourth image has the wrong arrangement.
- The bottom of the field, the open models that are a few years old, sit below 30 percent.
- On the subtasks that measure purely spatial relations the scores often fall below 30 percent.
- That is dangerously close to the 20 percent random baseline — in other words, on those subtasks the model is barely better than guessing.
It matters that these numbers are relative: they are good for comparing two models, not for predicting your particular image. We know anyway that image-model benchmarks drift away from human judgement — when one of the best-known metrics was reworked, the authors measured a 17.7 percentage point gap between the old score and the human ranking. So what you take from the numbers is that arrangement by words is unreliable, not exactly how unreliable.
Four methods, from the loosest to the strictest
There are four routes to specifying arrangement, and what separates them is not the effort but how much freedom you leave the model.
- Words. You describe the positions in the text of the prompt. The simplest, and sometimes enough — if only one thing needs placing and the tolerance is wide. Beyond two elements it starts to fall apart.
- Aspect ratio and empty space. You are not asking for a position but for a composition: "a large empty sky in the top third", "the subject low in the frame with room above it". This goes much better for models, because it is a compositional notion, not a coordinate.
- A sketch image. You hand over a drawing — a hand-drawn box layout or the outlines of an existing picture — and the model works from that. This is the most reliable method, and I will go into it in a moment.
- Structured description. You give the prompt not as a sentence but as a data structure, element by element with a style and optionally a box marking where it goes. One model family built its entire training around this, and there it is the native language.
The sketch as input
The point of the sketch-driven approach is that you hand over the structure of the image in picture form rather than in words. In the open-tool world there are mature control layers for this: the most common are edge detection, soft outline, depth map, body pose, freehand scribble and a greyscale base. By 2026 these have been folded into a single combined control model, so you no longer need a separate download for each.
In practice this means you open a drawing program and lay down the rectangles in two minutes: product here, caption space here, background horizon there. That becomes the sketch, and the model will hold that structure, because it does not have to interpret it, only follow it. On the kinds of reference image and what each of them carries over, the reference types post goes into more detail — the sketch is only one of them.
When the structure is the point
The structured description works on a different logic. You do not write a sentence, you list the elements, write each one's own style next to it, and you can state where it should sit. For a poster that is precisely what you need: the headline in this type at this size, the picture here, the small print there.
This is not a general capability today. One model family's training was built exclusively on descriptions with this structure, with per-element styling and optional bounding boxes — there it works, elsewhere it will most likely just turn into an unusually stilted prompt. If you regularly produce layout-dependent material, it is worth bringing this criterion into your choice of tool.
A recipe that is enough in most cases
If you do not want to wrestle with control layers, these three steps solve the arrangement for most marketing images.
1. Do not ask the picture for empty space, let the aspect ratio give it.
Ask for a tight crop on the subject and cut the space for the text
afterwards, or generate a wider image and extend its edges.
2. One arrangement request per image.
"The product in the left third" — that much. Do not add what goes in
the top right and what at the bottom, because that is where it breaks.
3. Do not let the model put the caption on.
Let the picture be a picture and the text come afterwards, in a
proper typeface. That is the only method that stays legible at
every size.
Common misconceptions
- "If I describe it in more detail, it will be more accurate." Not with arrangement. Several position requests in one prompt usually make the result worse, not better.
- "A better model will sort it out." According to the benchmarks the gap between the best and the middle of the field is smaller here than elsewhere — this is the weak point of the whole field.
- "A negative prompt helps keep it out of there." The negative prompt no longer even exists on several newer models, and it was never reliable for spatial requests.
- "Just regenerate enough times and it will hit it once." It will — but if that takes fifteen attempts, it is more expensive and slower than two minutes of sketching.
What to take away
Arrangement is the area where words run out soonest. If the composition genuinely matters — because text is going on top, because it goes into a brand template, because the product's position is fixed — then do not polish the prompt, give the model structure instead: a sketch, an aspect ratio or a structured description. The four layers that make up the rest of the prompt stay the same regardless, and the post on how an image prompt is built covers those in detail.