A target persona and an angle, chosen from evidence before a word is written.
for the frontier.
Enterprise-grade ad campaigns from one API call.
a creative director.
A model can generate an ad in seconds. Getting to one an enterprise brand will run takes a creative team and rounds of reprompts, because the first output has no persona, no angle, no job in the funnel, and a layout no creative director would approve. And a campaign isn't one ad. It's ten.
That's not a rendering problem. It's a data problem. Models were trained on all ads, not good ones, and the judgment that separates the two was never written down. It lives in the heads of a few thousand creative directors. We turned it into data.
The models got the hands. Nobody gave them the head.
from the first output.
“Your OB said no retinol for a year?”
“Skin is smoother. Fine lines less noticeable.”
20% off + free shipping, first order
Ten ads, each with a funnel job: stop a cold audience, make the case, close the sale.
Layouts our creative directors built and approved, matched to each ad’s job. Type set clean.
- Logo·✓
- Colors·✓
- Label·✗ Redraw✓
- Offer·✓
- Claims·✓
The brand’s real logo, colors, facts and live offers, on the real pack and label. Each ad is read back against the product and redrawn until it matches. Nothing generic, nothing guessed.
The head, for every model that already has the hands.
Dessert sits on top of whatever generates the ad: the frontier labs now selling ads, the platforms ads run on, and the tools they're made in. It makes the calls a creative director would make, the persona, the angle, each ad's job and the layout, and the model does the making. Whatever draws the pixels, the judgment comes from us.










One complete ad set
Ten finished ads.
Unedited.
Blind evals for advertising judgment. Frontier image models, Meta Advantage+, a top agency team and Dessert each got the same 40 real briefs and returned complete ten-ad sets. Six senior creative directors judged 2,000 ads without knowing who made them.
- 78%Ship rate
Ads creative directors would run unedited. Best frontier model: 17%.
- 89%Head-to-head
Sets preferred over the best frontier baseline, judged blind.
- 81%Grader agreement
Our taste judge agrees with the panel more often than its own members do.
- 7.9×Transfer
Ship rate of an open model trained on our decision trails, from 9% to 71%.
train on and measure.
Call it.
One tool call returns a complete ad set: the persona, the angle, ten ads with a funnel job each, laid out and checked against the real brand and product.
In private use with top DTC advertisersTrain on it.
Creative-director judgment as structured data: the evidence, the persona, each ad's job, the copy we rejected and why, every check it passed. Taste can be trained.
Licensable decision trailsMeasure with it.
A taste judge calibrated against blind creative-director panels. Use it to evaluate your models' creative, or as a reward signal to train them.
81% agreement with creative-director panels




