It is the axis on which models differ most and marketing says least. Two models can be equally capable of a beautiful frame while disagreeing completely about whether "holding the bottle in her left hand, label facing camera" is a requirement or a suggestion. For commercial work the second question is the one that costs money.
The hard cases are consistent across the field: counting (three of something, not four), spatial relations (behind, to the left of, on top of), text rendering, negation, and multiple attributes bound to multiple subjects — the red hat on the woman and the blue hat on the man, kept that way round.
Guidance settings raise adherence up to a point and then start distorting the image instead. Beyond that point the honest fix is either a different model or a different approach — reference images and masked edits pin down what a prompt keeps failing to.
In practice
- Test adherence with a brief that has a checkable requirement — a count, a position, a specific colour on a specific object.
- Put the non-negotiable requirement early in the prompt.
- If two attributes keep swapping between subjects, generate them separately and composite.
The mistake to avoid
Judging a model by its best output. Adherence is about the worst of five runs, because that is the one that shows what the model treats as optional.
Related terms
Prompt
A prompt is the written instruction a generative model reads to decide what to make — the one input almost every model requires.
CFG scale
CFG scale controls how strictly a model obeys your prompt, trading obedience against the model's own sense of what a natural image looks like.
Prompt expansion
Prompt expansion is a provider-side step that rewrites your short prompt into a longer, more detailed one before the generator ever sees it.
Negative prompt
A negative prompt lists what you do not want in the output, steering the generation away from those terms instead of towards them.
Temporal consistency
Temporal consistency is how well a generated clip keeps things the same from one frame to the next — a shirt that stays the same colour, a background that stays put.
The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync in a single app.