The short version
- A prior is what the model returns when your instruction is rarer in its training data than the alternative. Your prompt is one vote against the distribution.
- Forcing right-hand drive into a generated car interior took five attempts. Moving the camera so the contested detail sits outside the frame is the alternative to arguing for it.
- Every re-run pays the per-call overhead again. A one-image call measured 28.7k tokens and three images inside one call measured 38.7k, so batching is where that overhead gets amortised.
A model prior is the answer a generator defaults to when your instruction is rarer in its training data than the alternative. We spent five attempts forcing a right-hand-drive interior into a generated car for an Australian taxi payments company.
What is a model prior?
Ask an image model for a car interior and you get the interior the internet is full of, which means the wheel sits on the left. Nothing was misread. The model returns what it has seen most, and your prompt is one vote against millions of images.
Priors surface wherever the world is unevenly distributed, and steering wheels are only the obvious case.
Measured across our generation runs, fighting the model against steering it:
| What we wanted | Approach | Measured result |
|---|---|---|
| Right-hand-drive car interior | Force it in the prompt | 5 attempts |
| On-style ad creative | Style described in words | 2 of 3 usable |
| On-style ad creative | One style-reference image attached | 3 of 3 usable |
| Product screen inside a scene | Composite the screen onto a generated phone | 0 usable |
| Product screen inside a scene | Generate the screen into the scene in one pass | 3 of 3 usable |
Why does repeating the instruction not work?
Stronger wording changes the sentence you send. It does not change the distribution the image is sampled from.
The detail was not cosmetic, which is why we kept going. The set was running to an Australian audience, and the wheel was on the wrong side for that market.
What does fighting a prior cost?
Every generation call carries a fixed overhead of roughly 25,000 to 28,000 tokens before an image renders. A single-image call measured 28.7k in our runs.
Extra images inside one call cost about 5,000 tokens each, so three images measured 38.7k against 28.7k for one. Batching is where that overhead gets amortised, and every re-run pays it again from zero.
When do you change the composition instead?
Stop editing the prompt and edit the shot. Move the camera until the contested element sits outside the frame. Where the shot will not survive that, pick a scene where the prior and the requirement already agree.
The same account gave us the cleaner version of this lesson on a different shot. Compositing a product screen onto a generated phone produced nothing usable, and generating the screen into the scene in one pass produced 3 of 3.
You cannot argue a model out of its training data. Move the camera instead.
How do you find the priors before you burn five attempts?
Generate once from a plain prompt with none of your requirements in it. Whatever comes back unasked is the prior. Hold it against the brief and you know which requirements are free and which are a fight, before anyone commits to a shot list.
One generated ad set ran across the US, the UK and Australia, and those three markets do not agree on which side of the car the wheel belongs on. A frame that keeps the question out of shot serves all three from one render.
More on the same work: what has to be in the brief and why a rule you stated is not a rule you enforced.
Frequently asked questions
What is a model prior in AI image generation?
A prior is the output an image model defaults to when the thing you asked for appears less often in its training data than the alternative. Ask for a car interior and the wheel arrives on the left.
Why does an AI image generator ignore part of my prompt?
Because what you asked for is rarer in its training data than what it returns instead. Forcing a right-hand-drive interior into a generated car took five attempts.
What do you do when a generator keeps getting one detail wrong?
Change what the frame contains. Move the camera until the contested detail sits outside it. Arguing with the prompt again is what cost us five attempts on one steering wheel.
What does a failed AI image generation attempt cost?
Roughly 25,000 to 28,000 tokens of fixed overhead per call before anything renders. A single-image call measured 28.7k across our runs, and three images inside one call measured 38.7k.
How do you find a model's priors before starting a project?
Generate once from a plain prompt that contains none of your requirements. Whatever comes back unasked is the prior. Compare it against the brief and you know which requirements are free and which will be a fight.
Is it better to fight a prior or change the composition?
Change the composition. One generated ad set ran across the US, the UK and Australia, and those markets disagree about which side of the car the wheel belongs on. A frame that avoids the question serves all three.
Sources
- Chua Network delivery data across 8 client accounts (internal fact bank)
- Chua Network engagement records, anonymized (internal experience bank)
