Imagen
Why it matters
Google is one of the few companies with frontier models for text, images and video, and its image models reach an enormous audience through the Gemini app, Google Photos, Workspace, Search and Android. Imagen's research also shaped the whole field: its finding that the text encoder matters as much as the image model influenced Stable Diffusion 3, FLUX and many others.
Nano Banana, for its part, made conversational image editing mainstream. Uploading a photo and saying "make it evening and remove the car" while everything else, including people's faces, stayed consistent was a leap over earlier tools, and it drew hundreds of millions of image edits within weeks.
Deep Dive
Imagen was Google's text-to-image model family. The first version, published as a research paper in May 2022, showed that pairing a large language model as text encoder with a diffusion model produced images that followed prompts more faithfully than anything before it. Google held it back from public release at the time, then shipped successive versions in its products: Imagen 2 in late 2023, Imagen 3 in 2024 and Imagen 4 in 2025. Google has since shut down all Imagen models (Imagen 4 on 17 August 2026).
Alongside Imagen, Google built image generation directly into its Gemini models. That second line became famous in August 2025 under the nickname Nano Banana, and it now handles all of Google's image generation.
The Original Design
Imagen (2022) used a frozen T5-XXL text encoder and a cascade of diffusion models: one generated a 64×64 image, and two super-resolution models upscaled it to 256×256 and then 1024×1024. Unlike Stable Diffusion, it worked on pixels directly instead of in a compressed latent space. Later versions moved to more efficient latent diffusion designs, with steady gains in photorealism, text rendering and prompt following. Imagen 4 added sharper detail, much better typography, output up to 2K resolution and a fast variant.
Two Lines of Models
Until 2026, Google had two approaches to images:
- Imagen models were dedicated text-to-image generators. You sent a prompt and got images. They were fast, relatively cheap per image and strong on photorealism and style.
- Gemini image models generate images natively as part of a multimodal LLM. The first widely used one was Gemini 2.5 Flash Image (Nano Banana, August 2025). Nano Banana Pro, built on Gemini 3 Pro, followed in November 2025 with accurate text in images, output up to 4K, the ability to blend many reference images and optional grounding with Google Search. Further versions have continued the line since.
The practical difference is that the Gemini models understand the request with the full knowledge and reasoning of an LLM. They can edit over several conversational turns, keep a character consistent across images, follow complex layouts, draw diagrams and infographics with correct labels, and use world knowledge. For a single image from a clear prompt, Imagen was often enough. For editing, multi-image composition or anything that requires reasoning, the Gemini models were the better choice.
Where to Use Them
Consumers get the Gemini image models in the Gemini app and across Google products. Developers access them through the Gemini API in Google AI Studio and through Vertex AI on Google Cloud, priced per image or per output token. Third-party platforms offer them too.
Watermarks and Safeguards
Every image from Google's models carries SynthID, an invisible watermark embedded in the pixels and designed to survive cropping, compression and filters. Google lets users check images for it in the Gemini app, and it also attaches C2PA content credentials in some products. The models apply filters for unsafe content and limits on depicting real people, although the realism of recent versions has renewed concern about deepfakes and misinformation.
The Competitive Picture
The main alternatives are OpenAI's GPT Image models, which take a similar LLM-native approach, Midjourney for aesthetics, Black Forest Labs' FLUX for open weights and API use, and models from Chinese labs such as ByteDance's Seedream and Alibaba's Qwen-Image. The field moves quickly, and the lead on public leaderboards changes hands every few months.