Left, right, behind: the relations it cannot hold

The measurement, not the impression

Benchmarks built specifically for this — sets of prompts of the form "A to the left of B" with automatic checking — find that image models score far below what people assume from their general quality. Older models sat close to chance on left-and-right. Current ones are substantially better and still make errors at rates you would not tolerate in any other tool: a relation that fails one time in four is not a relation you can build a layout on.

The asymmetry is informative. Models handle above and below better than left and right. They handle on and in better than behind. And they handle relations between dissimilar objects better than between similar ones.

Each of those follows from the training captions. Vertical relations are described more often and more consistently, because gravity makes them stable and worth mentioning. Left and right depend on the viewer's frame and are frequently written from the subject's perspective instead, so the training signal is genuinely contradictory. Behind requires depth, which is not directly visible in a flat caption-image pair.

What "learned as association" means here

There is no coordinate system anywhere in the model. When it renders the cat to the left of the dog, it is not placing objects on a plane. Latent cells on the left side settle into cat-like content because the conditioning made cat slightly more likely in regions that co-occurred with left-ness in training. That is a soft statistical pressure competing with everything else in the image, and it loses whenever another pressure is stronger — for instance, the compositional habit of putting the larger subject in the centre.

This is the same machinery as attribute binding, and it fails for the same reason: no explicit structure holds the relation.

Solve it with geometry, not language

The reliable approach is to stop asking and start showing. Three routes, all free:

A conditioning map. Sketch two blobs in the right places in GIMP, Krita or any drawing app — twenty seconds, no skill required. Feed it as a scribble or depth control. The layout is then fixed by the map, and the prompt only has to say what the blobs are.

Regional prompting. Divide the canvas explicitly and condition each region. This solves relation and binding at once, since the left region's prompt contains only the cat.

Composite. Generate the cat, generate the dog, place them. This is how nearly all commercial work is done anyway, and it also solves scale, which relational prompts never handle: a cat next to a dog gives you no control over their relative size, and the model's default is frequently wrong.

The related failure with viewpoint

A neighbouring problem worth naming, since it wastes as much time. Prompts specifying camera position — low angle, from behind, three-quarter view, bird's eye — work only to the degree those phrases appeared in captions with consistent meaning. Low angle and bird's eye are photographic vocabulary and work reasonably. From behind and slightly to the left is not a caption anyone writes and behaves accordingly.

For anything where viewpoint matters, the cheap professional trick is to build the scene roughly in free 3D software — Blender, or even a simple posing tool — screenshot it, and use the screenshot as a depth or edge map. You get exact camera control from a program designed for exact camera control, and the generative model does what it is good at, which is surfaces.

The honest limitation: none of these give you a model that understands space. They give you ways to supply the spatial information yourself. That distinction matters when someone claims a new model has solved composition. Ask how it does on relations between two similar objects, and on "behind". Those are the cases where the association is weakest and the claim gets tested.

A last practical note, because it explains a lot of wasted time. Relation failures are not stable across seeds, which means the first thing everybody does — regenerate and hope — sometimes works. That is the worst possible feedback. Getting the relation right one time in three teaches you that the prompt nearly works and encourages another twenty attempts, when the actual state of affairs is a coin weighted slightly in your favour. The test for whether a relational prompt is reliable is four fixed seeds, not one lucky one. If it fails on two of four, it will fail on the shot the client picks.