AI&Design
All change articles

Gemini image · Announcement · Official source

What early Imagen 4 users learned about lettering and layout

On May 20, 2025, a Gemini user shared two posters for an imaginary music festival on Europa and praised the lettering. The prompt named the festival, date, and location. Later in the thread, the author identified the Gemini app and another reader asked how they could tell which image model had run. The poster remains a record of what the author says the app produced. Its exact model routing is unconfirmed. The festival posters and prompt

A second user that day reused a prompt for an ancient Roman street scene, with a Fujifilm camera and lens specified as photographic cues. The author reported that all four outputs put a camera in the scene and also complained about faces and fine details. The two posts expose different demands on a prompt. A festival name belongs in the picture as literal text. A camera name can describe how the picture should look. The street-scene trial

What Google changed in Gemini

Gemini already offered image generation with Imagen 3. In its May 20 release notes, Google introduced Imagen 4 in the app and singled out spelling and typography, with posters and invitations among the intended uses. That gave users an explicit reason to put copy into a visual brief. The release claim addresses lettering. The surrounding composition and the meaning of a phrase like a camera specification still have to be evaluated in the resulting image. Imagen 3 in Gemini · Gemini Apps release notes · Google I/O launch

Where the brief held up

In a June 22 report, Mitja Rutnik described getting requested aspect ratios in the Gemini app, including 16:9 and 9:16. A jacket recolor gave him a different kind of test. After he asked Gemini to keep the rest of the image the same, the result carried over much of the subject and setting, with a different pose and camera angle. For someone laying out a poster, the requested canvas shape helped. For someone revising a selected image, the change in viewpoint could undo a layout choice. The report documents the app's behavior in that session. The image model and configuration for each turn remain unconfirmed. Rutnik's format and revision trials

Blake Harkness tested a denser lettering problem in a report published July 5. His brief called for an infographic about the New Zealand housing market. He found the layout of the output he attributed to Imagen 4 promising at first glance, then found garbled headings and small text and unusable pie-chart values. The music poster and the housing graphic illustrate different text burdens. The former asks for a handful of supplied strings. The latter requires the reader to be able to trust labels and figures throughout the design. Harkness's housing infographic trial

Where Imagen 4 fit in a cover workflow

The developer route had its own release dates. Google announced paid preview in the Gemini API and limited testing in AI Studio on June 24, then general availability on August 15. In a September 7 AI Studio challenge write-up, a builder posting as Ha3k described using imagen-4.0-generate-001 to generate initial book-cover concepts at a fixed 9:16 ratio. A selected image and the edit instruction then went to Gemini 2.5 Flash Image Preview. The prototype gives the generation step an explicit model and format. The revision step has its own model, and publication of a finished cover remains unconfirmed. June developer preview · August general availability · CoverCanvas build write-up

These accounts show how an image can meet the brief on aspect ratio, headline, or color and still need review. A follow-up test could carry the same brief through a first render and a targeted revision, recording the model and product surface at each turn. The brief would determine which copy must be exact, which chart values must be faithful to the source, and which parts of the composition must stay in place.