Google releases Imagen 3.5: Google's best text-to-image model

Generating images with AI is easy. Generating them with correctly spelled text and the right proportions is another story. Google's Imagen 3.5 solves exactly this: it's the first Google model that handles text in images in a truly reliable way. Write "Welcome to the Workshop" and the image shows exactly that, not "Welcme to the Workshp".

The photographic realism has improved significantly compared to Imagen 3: skin, fabrics, natural light look like real photos. For content creators, designers, and marketing teams this is a concrete advantage.

The most important new feature on the responsibility front is SynthID enabled by default: every generated image has an invisible watermark certifying its AI origin. It's not visible to the naked eye, not easily removable, and Google can verify it. In Gemini you can already use it today to generate images in conversations. On Vertex AI it's available via API with per-image pricing. In Workspace, Slides and Docs can generate images directly in the document. For those already using the Google ecosystem, it's a transparent and immediate upgrade.