The most common complaint from brand teams experimenting with generative production is not quality. It is drift. Image one is excellent. Image twelve is a different brand.
Why drift happens
Generative models optimise for plausibility, not for repetition. Every run is a fresh sample. Without something anchoring it, the output will wander toward the average of everything the model has seen — which is, by definition, not your brand.
What actually holds a look together
- A reference set, not a prompt. A curated, deliberately small body of approved imagery that defines the target. Prompts drift; references anchor.
- Fixed parameters. Lens behaviour, colour temperature, contrast curve, depth of field, grain. Written down and reused, not rediscovered.
- Composition rules. Where the subject sits, how much air, what the horizon does. Layout is half of brand recognition and it is fully controllable.
- A fixed post treatment. One grade, one grain, one finishing pass applied to everything. This alone removes most visible drift.
- Built elements. Product, type and logo composited rather than generated, so the parts that must be identical always are.
A brand look is a set of constraints. Generative tools have no opinion — so you have to supply one.
Documenting it so it survives
A system that only works when one person runs it is not a system. The deliverable is not a folder of images — it is the reference set, the parameters, the post recipe and the rules, written so an in-house team can produce asset two hundred without asking anyone.
The honest limitation
Even a good system will not deliver frame-exact repetition. It delivers family resemblance — which is what brand consistency has always actually been. Where exact repetition is required, that element gets built and composited. Knowing which elements those are is the whole design of the system.