Opinion
Creative AI needs to know what not to change
"Advertising production reopens some creative decisions," writes Bria CTO Michael (Misha) Feinstein. "AI has to preserve the rest."
An approved ad is not a blank canvas. By the time it reaches production, decisions about the product, logo, composition, typography, colors and hierarchy have already been made. The production challenge is to adapt the asset without reopening all of them.
Most production requests touch only part of the creative. Localization can force the layout to rebalance, and a new format can change the composition, while the product, logo and visual hierarchy remain approved. The job is to change what the brief reopened and carry the rest forward.
I often think about the difference between a casual user and a film director. A casual user can be delighted by a beautiful surprise. A director often already has the frame in mind. If the system produces something impressive but different, the result is still wrong for the job. Once a creative direction is approved, the next edit has to respect the decisions that approval closed.
In engineering terms, production requires state: which decisions are fixed, which are variable, and what the current instruction is allowed to affect. A prompt describes a target; by itself it does not preserve the state of a working file. The next request is usually smaller: change this line of copy, keep the product untouched, move the offer, recompose for a new placement, preserve the brand color. The useful unit is often the change itself, not a new image.
A layered creative file already preserves part of that state. Text, logos, products and backgrounds remain separate enough for a designer to change one without rebuilding the rest. AI needs to carry that structure forward instead of collapsing everything back into a final image.
At Bria, this principle has shaped our approach to structured visual representations and editable ad layers. A finished flat ad is reconstructed into editable elements, so the next request starts from retained structure instead of a new generation. Copy remains editable as copy; a background edit leaves the product intact.
Localization makes the boundary more demanding. Translated copy takes more or less space than the original. Hebrew or Arabic changes reading direction. A new placement forces the composition to rebalance. Preserving intent still allows the layout to move. The product, logo or an approved brand treatment may need to remain exact.
Visual agents raise the stakes because they handle sequences ratherthan isolated edits. An agent turning one campaign into dozens of assets makes many small production decisions along the way. If every result returns as a flattened image that a designer has to rebuild, automation has only moved the bottleneck. People set the creative intent and the boundaries; the system executes the repetitive steps inside them and returns editable material when judgment is needed.
In engineering, I do not call something done because the model works. It has to work in production, where it is monitored, measured and relied on. A strong first image says little about that standard. Production quality shows up in repeated changes, corrections, handoffs and reuse.
A production test starts with an approved asset. Localize it, resize it, change one element and preserve another. Track what drifted, what survived, and how much manual rework remained before the asset was ready. The useful result is how much approved creative survives the transformation and how much repair remains.
As image quality rises, the difference between a demo and a production system moves out of the screenshot and into the life of the campaign. The first creative concentrates human judgment. The assets that follow carry that judgment across formats, languages and placements.
AI earns a real role in production when each edit reopens only the decisions the brief actually changed.
Michael (Misha) Feinstein is CTO at Bria.
