Imagen 2

A second-generation model from Google

Imagen 2 is the second-generation text-to-image model from Google, released in December 2023.^[1][2] A text-to-image model takes a written prompt — a sentence describing a scene, an object or a style — and returns an image meant to match it. The first generation of such models answered the question of whether the approach worked at all; the second generation is usually where the obvious failures get attention, and for Imagen 2 that failure was writing.

Legible text and logos

The standout feature of Imagen 2 was the generation of text and logos within images.^[1][2] Google Cloud said text-to-image models often have difficulty rendering text correctly, and that Imagen 2 helps solve this.^[2] The difficulty is easy to see. Letters are thin strokes, tight spacing and exact shapes, and a model that is busy getting the composition, lighting and colour of a picture right can easily produce lettering that looks plausible from a distance and turns into nonsense when read. To fix that, the model has to hold on to small, precise detail while the rest of the picture is being made — a different demand from producing a convincing scene. It matters in practice too: posters, advertisements and product mock-ups almost always carry words, and a picture whose words are garbled is hard to use.

Logos were the other half of the same feature. Imagen 2 can generate logos, including emblems, lettermarks and abstract logos, for businesses, brands and products.^[2] These are standard categories in logo design: an emblem wraps a name or motto into a symbol or picture; a lettermark builds the mark out of the initials or letters of a name; an abstract logo uses a geometric or otherwise non-representational form rather than a picture of a thing. A generated logo is a harder target than a generated illustration, because a mark is judged on whether it stays consistent and recognisable across many uses, not on whether it looks good once.

Available on Vertex AI

Google Cloud made Imagen 2 generally available for Vertex AI customers on the allowlist in December 2023.^[2] On Vertex AI, customers could customize and deploy the model with fully-managed infrastructure and built-in privacy and safety features.^[2] That is roughly the difference between a demonstration and a product: a business could adapt the model and run it inside Google Cloud's managed environment, with privacy and safety controls supplied by the platform, instead of assembling that machinery itself.

Its place in the line

Imagen 2 was succeeded by Imagen 3, which Google announced on 14 May 2024 and released in August 2024.^[1][3] Google described Imagen 3 as its highest-quality text-to-image model, producing photorealistic images with far fewer distracting visual artifacts than its prior models.^[3] The two releases are less than a year apart, and the emphasis moves from a specific defect — text that would not render — to overall image quality and the removal of visual artifacts.