Imagen 3

Imagen 3 is a text-to-image model from Google. You describe a scene in words, and the model produces an image. Google announced it at its I/O developer conference on 14 May 2024, and described it as the company's highest-quality text-to-image model, with fewer visual artifacts and improved text rendering.^[1]

What it does

A text-to-image model takes a written prompt and returns a picture. Work in this area is usually judged on two things. The first is faithfulness: does the picture contain the objects, the arrangement and the style the prompt asked for. The second is execution: do small details — hands, edges, lettering, background objects — hold up when you look closely. Visual artifacts is a general term for failures of the second kind, such as smeared edges, malformed fingers, or objects that melt into one another. Google's claim for Imagen 3 was that it produced fewer of these, and that it was better at rendering text inside an image.^[1]

Rollout

Imagen 3 was not available to everyone at once. At launch it was in private preview for select creators in ImageFX, Google's image generation tool, through a waitlist, and Google said Vertex AI would follow soon.^[1] The same announcement repeated that the model would come soon to Vertex AI, the platform developers use to run Google's models.^[1] The release itself came in August 2024.^[2]

Access widened later. In December 2024 the latest Imagen 3 rolled out in ImageFX to more than 100 countries.^[3] Around the same time Google launched Whisk, a Labs experiment that combines Imagen 3 with Gemini's visual understanding — the side of Gemini that reads images rather than producing them.^[3]

For developers, the model was offered through the Gemini API under the identifier imagen-3.0-generate-002, released on 6 February 2025.^[4] That version did not stay long: the same identifier was shut down on 10 November 2025, and the API's deprecation page recommends gemini-nano-banana-2.1 as the replacement.^[4]

Text rendering

Lettering is a particular weak point of image generation, and it is easy to see why. A word is a sequence of shapes with exact outlines; a stroke in the wrong place turns one letter into another, and a stretched or invented character is immediately obvious to any reader. Models that learn from whole pictures are not naturally organised around that kind of character-level structure. Google framed text rendering as a known challenge for image generation models and said Imagen 3 was its best model yet at it.^[1] Improvements in this area matter beyond posters and signs: prompts often ask for written words on objects, and a model that garbles them produces images that cannot be used as they are.

Where it sits

Imagen 3 belongs to Google's Imagen line of image models, and the way it was distributed shows a pattern that has become common for such models: a consumer-facing tool on one side, ImageFX, and a developer route on the other, through Vertex AI and the Gemini API.^[1][3][4] Model versions are released and retired on their own schedule, independently of the product they sit in — the Gemini API identifier for Imagen 3 had a published start and end date, and a named successor.^[4]

Pairing it with Gemini in Whisk points in another direction: using a generation model together with a model that can look at images and understand them, rather than treating generation as a standalone task.^[3] At launch, though, Google's own framing was narrower — quality and text rendering, rather than a new kind of capability.^[1]