Prompt adherence meaning

Adherence is what separates a model that treats 'three red apples on the left' as a requirement from one that treats it as a mood. The count, the positions, the colours, the text and the action are the details that show it, and you measure it by writing a prompt with details you can check, then counting how many a model gets right across several runs.

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. Research benchmarks score it the same way, with checkable prompts. GenEval (2023) runs an object detector over generated images to check object co-occurrence, position, count and colour, and found models weakest at spatial relations and attribute binding. T2I-CompBench++ uses 8,000 compositional prompts in four groups: attribute binding, object relationships, generative numeracy and complex compositions. A published score is a fair first filter; your own brief, with your own checkable details, is the test that decides. Usable rate is the same idea applied to a whole job.

Guidance settings such as CFG scale 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.