Hands, faces and other hard things, in order of how fixable they are

The default response to a bad generation is to run it again. Sometimes that works, which is what makes it such a durable habit — and why people burn afternoons re-rolling something that was never going to come out.

Failures fall into three classes with different economics. Class one is fixed by wording. Class two is fixed by changing what you asked for. Class three is not currently fixable and the correct response is to redesign the shot.

Knowing which class you are in is the whole skill. The list below is ordered by how tractable each problem is.

Flat, shapeless lighting. The default output of most models. Fixed by naming a single source and its direction. Four words, reliable.

Centred, symmetrical composition. Also a default. Fixed by stating where the subject sits — off-centre, in the upper third, with the lower area empty.

Too much in the frame. Models add objects to fill space. Fixed by an explicit exclusion: no additional objects enter the frame.

Plastic skin. Fixed by naming a film stock and asking for visible pores and texture, rather than by any amount of quality vocabulary.

All four are among the most common complaints and all four are prompt problems. Try the wording change before the re-roll.

Hands. Much improved and still unreliable at scale. A hand occupying a small part of the frame is usually fine; a hand as the subject, doing something specific, is a gamble. Fix by reframing — crop the hand out, obscure it, or make it a supporting element rather than the focus.

Interacting objects. Two things touching, gripping or overlapping is markedly harder than two things near each other. Fix by separating them in the composition.

Reflections showing a scene. A mirror or glossy surface reflecting the room is reconstructed rather than computed, and it will be wrong. Fix by choosing matte surfaces or angling reflective ones away.

Crowds. Faces in the background degrade quickly. Fix by putting distance, blur or backs to camera between the viewer and the crowd.

In every case, the successful move is to redesign the shot rather than to insist on it. This is what a director does when a location will not give them the angle.

Small readable text. Covered elsewhere, but it belongs on this list. Set it afterwards.

A specific real person. Beyond the ethical question, models do not reproduce a particular individual reliably from description, and approximating a recognisable person is a problem in itself.

Exact brand assets. A logo, a specific typeface, a defined packaging layout. Composite the real file.

Anatomically precise work. Anything where correctness is checkable against a standard — medical, mechanical, architectural detail that has to be right — will look plausible and be wrong.

Consistent text across a sequence. Even where a single frame spells correctly, the next frame will not match it.

1. Is the failure in class three? If so, stop and redesign the shot. No further attempts.
2. Is it composition or lighting? Change that block of the prompt and regenerate once.
3. Is it a specific object being wrong? Remove it from the prompt entirely and see whether the rest holds. If it does, the object needs to be composited or reframed.
4. Is everything nearly right? Now re-roll — this is the one case where the seed is genuinely the variable.

Most people start at step four. Starting at step one saves the afternoon.

These categories move. Things that were class three two years ago are class one now, and a new model can shift several at once — usually without announcing which.

Keep a short file of failures you hit and what resolved them. Re-test the class three list whenever you change models; occasionally something has quietly become possible and you would otherwise keep designing around a limit that no longer exists.

The categories themselves are stable. What moves is which items sit in each one, and that is worth checking rather than assuming.