Why image models misspell
An image model draws pixels, and to it a letter is a shape like any other. Newer models get short words right more often, but long words, small type and anything drawn twice still change: a headline that is right in one drawing loses a letter in the next. For an ad, where every word is checked by the brand and read by the buyer, "usually right" is not good enough.
Split the picture and the words
Hyzgen's static ads are two layers. The image model draws only the picture, with no text of its own. The headline, the button, the badge and the logo are then set over it by a layout engine, the way a designer would set them in a design tool. Hyzgen uses Satori, an open-source renderer that turns a layout into vector graphics, and runs the same layout in the browser for the preview and on the server for the download, so what you see is what you get.
Splitting them has three side effects that matter as much as spelling. Changing a word costs nothing, because nothing is redrawn. Other sizes are a new layout of the same picture, not a new drawing. And the words can be in any font the brand uses.
Keep the product's own lettering
A blanket "no text" rule keeps the picture clean and erases the product with it: handed an app screenshot under that rule, a model returned a blank device. For product scenes, say instead that the model adds no text of its own, and the only lettering is what is already on the reference.
Screens are the hardest case. A model asked to show an app invents its screen, text and numbers included. Hyzgen asks for a phone with a plain green screen, finds that green area in the picture, and sets the real screenshot into it in perspective. One detail: writing the colour as a hex code in the prompt got the code painted on the screen, so the prompt names it in words.
Leave room for the words
Draw the picture at the shape of the space it will fill, not the whole ad, and say where the words will sit so the product stays clear of them. Models follow that loosely: a phone can still end up under the headline. So let people move the picture's focus point and zoom after it is drawn, without drawing it again.
Fit and case the words
Nothing in a layout engine like Satori measures text for you, so a headline has to be sized by estimating how it wraps, from each font's average letter width. Long copy should end on a whole sentence: cut on a word, one slide read "...better than a cluttered".
Writing models love Title Case and capitals, whatever the prompt says. Hyzgen lowers a line that is mostly capitals to sentence case, keeping the brand's own names, names from the brand's About text, and codes like EID20.
Logos that stay visible
A logo set straight on a photo disappears when their colours meet: a white logo vanished on a bright picture. Put it on a small chip. When a logo has to change colour for a dark or light ground, recolour its type rather than inverting it: a plain invert turned a lime mark dark green.
Keep the copy honest
- Offers only when the brief states them. A discount the brand did not offer is a promise someone has to honour.
- Prices only from the catalogue, never from the writing model.
- No example claims in the prompt. "New colour" was given as an example badge, and the model put it on an ad.
- No "Try it free" by habit. It was in the examples, and it appeared on ads for products with no free plan.
Keep the ad's own text fields short too. Meta's ads guide suggests 50 to 150 characters of primary text and a headline of about 27 characters.
Right-to-left headlines
Most layout engines shape Arabic-script letters but still lay the words out left to right. An Urdu headline came out with its words backwards and its dots adrift. Hyzgen therefore shapes Urdu headlines itself with HarfBuzz, orders the words with the Unicode bidirectional algorithm, and hands the layout finished outlines. If your tool sets Urdu or Arabic, check a headline that mixes in an English word and a number.