The model spells the words wrong
Image generators draw text into the picture, and you get a missing letter or a smear where the headline should be. You notice once it is already in the ad account.
Give the studio a landing page URL or a short brief and it builds banners at the exact sizes Meta, Google, TikTok, Telegram and YouTube expect. A vision model then looks at the finished pixels — rendered text, artefacts and ad safety included.
Image generators draw text into the picture, and you get a missing letter or a smear where the headline should be. You notice once it is already in the ad account.
A square instead of 1080×1350, a cropped logo, copy outside the safe area. The platform crops the layout and the impressions go to a mangled creative.
English copy on a banner aimed at an Arabic or Hebrew market breaks both meaning and layout: the writing direction is the other way round, the design is not.
The studio reads the site and the client's market context, and pulls out the offer and the audience. A short written brief works just as well when there is no site yet.
The model renders a text-free image, the copywriter writes in the campaign language, and the composer lays that copy onto the layout — right to left for Hebrew and Arabic.
A vision model inspects the pixels while the audit compares the layout to the platform spec. What passes is exported in every size, or turned into an MP4.
A size catalogue for Meta, Google, TikTok, Telegram and YouTube: from 1080×1080 and 1080×1920 to 728×90 and 300×250, with the recommended ones marked.
Ad copy is written natively in Russian, Ukrainian, English, Spanish, Hebrew or Arabic and checked for script, length and unsupported claims.
The vision model reads the text rendered in the image and looks for anatomy defects, watermarks and artefacts. With no provider key the review reports «unavailable», never «passed».
Safe area, share of visible text and mandatory checks give each creative a status: launch ready, needs review, or blocked.
ffmpeg turns a finished banner into a Ken Burns move — a zoom or a pan cut to the exact placement size. No paid provider is involved.
Gemini and ChatGPT packages with image and video slots and included regenerations are optional add-ons; the base mode works without either.
Yes, that is the main Lite path: you give a URL, the studio reads the page, extracts the offer, tone and audience, and assembles a set of banners for the platforms you picked. A written brief works too when there is no site yet.
The image is generated without any lettering and the copy is added as a separate layer in the browser, so the words are always correct. On top of that a vision model inspects the finished pixels and reports any stray text the generator still painted in.
Russian, Ukrainian, English, Spanish, Hebrew and Arabic. For Hebrew and Arabic the layout is composed right to left, and Noto Sans Hebrew and Noto Sans Arabic sit in the font stack so the letters do not fall apart.
Not for the Ken Burns move: that clip is rendered by ffmpeg on your own server and costs nothing. Generative video from Veo or Sora runs through the Gemini and ChatGPT premium packages, where slots are bought separately.
It gets a «needs review» or «blocked» status with the concrete complaints listed — from copy outside the safe area to artefacts in the image. You can regenerate the banner without rewriting the brief from scratch.
A landing page URL is enough: correct formats, copy in the market's language and a pixel check are on from the first run.