Prompt adherence is how literally a model follows what you asked for, as opposed to producing something impressive in the general vicinity of it.
It is the property that separates a model you can direct from a model you negotiate with, and for production work it matters more than raw image quality. A beautiful shot that is not the shot in the brief costs you a regeneration; a plainer shot that is exactly right ships.
What good adherence looks like
You ask for a slow push-in at eye level, warm single-source light from camera left, and the subject glancing away at the end. A model with strong adherence gives you those four things. A model with weak adherence gives you a beautiful shot with a drifting camera, even lighting, and a subject staring down the lens.
The tell is not whether the output is good. It is whether the specific clauses survived.
Why it matters more than quality
Because quality is now broadly high across the leading models while adherence varies a lot, and because adherence determines your attempt count. A model that lands your prompt in two tries at 86 credits costs 172 credits per usable clip; a prettier model that lands it in five at 54 costs 270 and takes longer. See how to choose an AI video model.
Adherence is also what makes a shot list actionable. If the model ignores half the clauses, the shot list becomes a wish list.
How to test it in ten minutes
Write one prompt with four separately checkable clauses: a camera move, a light direction, a specific action, and a duration or framing. Run it three times on each candidate model. Then score how many of the four clauses appear, per generation, rather than judging the output as a whole.
That score is adherence. It correlates with how the model behaves on your real work far better than any leaderboard does, because it is measured on your prompt.
Where adherence breaks, in every model
Compound instructions. Four things happening in sequence within one clip. Models handle state changes over time poorly.
Negations. "No text on screen" frequently produces text. Describe what should be there instead.
Precise counts. Three objects tends to produce two or four.
Text inside the image. Improving, still the least reliable area.
Related concepts
What is negative prompting covers the technique for excluding elements. How to make AI video look less like AI covers prompt clauses that remove model defaults. Best AI video generator 2026 compares the field on identical prompts.
Score the clauses, not the beauty. The 8frame canvas is free and unlimited, and generation is paid from $19/month.