Imagen 3

Imagen 3 is Google DeepMind's third-generation text-to-image model, unveiled at Google I/O in May 2024 and released broadly in August 2024. It generates detailed, photorealistic images from plain natural-language prompts, and was widely noted as the strongest model of its generation at rendering legible text inside images — a long-standing weakness of diffusion models.

Under the hood it combines transformer-based language models (in the T5 lineage) for prompt understanding with cascaded diffusion for high-fidelity synthesis, and every output carries Google's SynthID invisible watermark. Human evaluations at launch put it at or near state-of-the-art against the leading commercial image models.

Imagen 3 is closed and API/platform-only: it is served through Gemini, ImageFX, and Vertex AI for enterprise customers. No weights are published, and access requires a Google account or a Google Cloud project.