Why does the model ignore the camera angle I asked for?

Because camera position and subject pose compete for the same part of the description, and the subject usually wins. A model has seen far more captions describing people than describing lenses, so a phrase like low angle is often applied to the body rather than to the viewpoint. Naming the viewpoint as its own clause, separately from the subject, is what separates them.

How do I describe a pose so it actually follows?

In terms of where the body parts are, not in terms of the mood. Arms crossed, weight on the left leg, head turned three quarters to the right is followed far more reliably than confident or relaxed, because the first describes geometry the model can render and the second describes an impression it has to infer.

Guide

Getting the pose and the camera angle you asked for

Ask for a low angle and the subject crouches. Ask for a side view and the head turns but the camera does not move. These are not random failures — the two instructions are competing, and one of them reliably wins.

The confusion

Where the camera is, and where the body is

A picture of a person contains two independent facts: how the body is arranged, and where the lens was standing. A human reader separates them without effort. A model does not, because in the captions it learned from they were almost never separated.

Most photo captions describe the subject. A much smaller number describe the viewpoint, and when they do it is usually in the same breath — “low angle shot of a man”. So the association between the words and the geometry is weak, while the association between the words and the subject is strong. Asked for a low angle, the likeliest thing the model has seen is a person doing something low.

This is the same mechanism as a prompt being half ignored, applied to one particular pair of instructions: the weaker claim loses to the stronger one rather than both being honoured.

Separating them

Two clauses, subject first, camera second

The reliable structure is to finish describing the subject, stop, and then describe the camera as its own statement about the photograph rather than about the person.

Instead of low angle photo of a woman standing in a kitchen, which tends to put her on the floor.

Write a woman standing in a kitchen, weight on her left leg, looking out of frame to the right. The camera is low, near waist height, tilted slightly up.

The second version costs more words and gets the geometry. Putting the camera clause last matters too: the end of a prompt is weaker on attention generally, but a short clause that introduces no new subject survives there, where a competing subject would not.

Pose

Describe the geometry, not the mood

Confident, relaxed, elegant and dynamic are the four most common pose words and the four least effective. Each describes an impression the model has to work backwards from, and the route from impression to limb position is exactly where it will substitute whatever was most common in training.

Body language is geometry, so write geometry. Arms crossed, chin lowered, weight on the back foot, head turned three quarters to the right, eyes to camera is a set of positions that can be rendered one at a time. It also fails more gracefully: if one clause is dropped you lose a detail rather than the whole arrangement.

Hands deserve their own clause or none at all. A hand given nothing to do is a hand the model will invent something for, and invention is where hands fail. In a pocket, at the side, holding one named object — or out of frame.

When words are not enough

Some geometry has to be shown

There is a ceiling here and it is worth knowing where it is. A specific pose — this exact arrangement of limbs, repeated — is not reliably reachable through description at all, however carefully written. Language is a lossy way to specify geometry, and two readers given the same paragraph would not stand the same way either.

Where the pose genuinely matters, the answer is to supply it as a picture rather than a paragraph: a reference image carries the arrangement directly. Words are for the camera, the light and the scene; images are for the shapes that have to be exact.

Asked often

Short answers

Why does the model ignore the camera angle I asked for?: Because camera position and subject pose compete for the same part of the description, and the subject usually wins. A model has seen far more captions describing people than describing lenses, so a phrase like low angle is often applied to the body rather than to the viewpoint. Naming the viewpoint as its own clause, separately from the subject, is what separates them.
How do I describe a pose so it actually follows?: In terms of where the body parts are, not in terms of the mood. Arms crossed, weight on the left leg, head turned three quarters to the right is followed far more reliably than confident or relaxed, because the first describes geometry the model can render and the second describes an impression it has to infer.

Keep reading

The rest of the series

Where this comes from

Unlimited Studio is built around this one problem

Everything above is what the product does: a character is locked from eight reference angles, each generation conditions on whichever angle its prompt implies, a score checks the result against the reference, and a film is shot one frame-chained shot at a time. What it will and will not generate is set out in the content policy.

How the pipeline works What it costs