Saharia et al. at Google Research published Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding (Imagen), proving that conditioning cascaded Diffusion Models on embeddings from a large frozen T5-XXL text encoder dramatically improved compositional prompt adherence—a design subsequently adopted by SD3 and FLUX.1.
Imagen
Also known as: Google Imagen · Imagen 2 · Imagen 3 · Imagen 4
Plain English
Google DeepMind's family of text-to-image diffusion models that pioneered using large frozen language models for prompt understanding and powers image generation across Gemini, ImageFX, and Vertex AI.
Technical Overview
The original 2022 Imagen architecture showed that scaling a frozen pretrained language encoder (T5-XXL) yields larger gains in text-image alignment than scaling the diffusion backbone alone. Later generations (Imagen 3 and Imagen 4) migrated to high-resolution latent diffusion trained with detailed synthetic recaptioning for photorealism and embedded typography.
Timeline (4)
2022
Paper2023
ReleaseGoogle DeepMind and Google Cloud launched Imagen 2 on Vertex AI, introducing multilingual text-overlay rendering, logo generation, inpainting/outpainting APIs, and invisible SynthID watermarking.
2024
PaperGoogle DeepMind published the Imagen 3 technical report and rolled out the model across ImageFX and Vertex AI, leveraging rich multi-level synthetic recaptioning to reduce prompt-binding errors and improve fine lighting and text rendering.
2025
ReleaseGoogle DeepMind introduced Imagen 4 and Imagen 4 Ultra at Google I/O 2025, scaling native generation up to 2K resolution with sharper fine fabric/water textures and reliable layout typography for slides, posters, and comics.
See also (9)
Open full graphImagen