Image 3 – Google DeepMind

Imagen 3 is our highest quality text-to-image model, capable of generating images with even better detail, richer lighting and fewer distracting artifacts than our previous models.

We’ve significantly improved Imagen 3’s ability to understand prompts, which helps models generate a wide range of visual styles and capture small details from longer prompts.

To be even more useful, Imagen 3 will be available in multiple versions, each optimized for different types of tasks, from generating quick sketches to high-resolution images.

Starting today, Imagen 3 is available to select creators as a private preview in ImageFX, and you can sign up to join the waitlist. Imagen 3 will also be available soon on Vertex AI.

Greater versatility and quick understanding

We designed Imagen 3 to generate high-quality images in a wide range of formats and styles, from photorealistic landscapes to richly textured oil paintings or fanciful clay scenes.

Imagen 3 also includes prompts written in natural, everyday language, making it easy to achieve the desired result without complex prompt engineering.

To help Imagen 3 capture nuances like specific camera angles or compositions in long, complex prompts, we added richer details to each image’s caption in its training data. With better information to learn from, Imagen 3 more accurately generates a wide range of topics and styles.

Better quality images

Imagen 3 generates visually rich, high-quality images with good lighting and composition. It can accurately render small details like the fine wrinkles on a person’s hand and complex textures like a knitted stuffed elephant.

Better rendering of text

We’ve also significantly improved its text rendering capabilities, opening up new usage possibilities such as stylized birthday cards, presentations and much more.

Built and deployed with our latest security and accountability innovations

Imagen 3 was built with our latest security and accountability innovations, from data and model development to production.

We used extensive data filtering and labeling to minimize harmful content in the datasets and reduce the likelihood of harmful results. We also conducted red teams and reviews on topics like fairness, bias, and content safety.

We deploy Imagen 3 with our latest privacy, safety and security technologies, including our innovative watermarking tool SynthID, which embeds a digital watermark directly into the pixels of the image, making it detectable for identification but imperceptible to the human eye.

Over the coming months, we will make popular Imagen 2 editing features, such as inpainting and outpainting, available in Imagen 3. And we will expand the availability of Imagen 3 to all Google products, such as the Gemini app and web experience, Workspace, Ads and more.

Thanks

Main contributors

Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Sergio Gómez Colmenarejo, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, Hongliang Fei, Yilin Gao, Evgeny Gladchenko, Mandy Guo, Alex Haig, Will Hawkins , Hexiang (Frank) Hu, Huilian Huang, Tobenna Peter Igwe, Christos Kaplanis, Siavash Khodadadeh, Ksenia Konyushkova, Karol Langner, Eric Lau, Shixin Luo, Soňa Mokrá, Henna Nandwani, Yasumasa Onoe, Aäron van den Oord, Zarana Parekh, Jordi Pont-Tuset, Hang Qi, Rui Qian, Deepak Ramachandran, Robert Riachi, Hansa Srinivasan, Srivatsan Srinivasan, Robin Strudel, Benigno Uria, Oliver Wang, Su Wang, Austin Waters, Chris Wolff, Auriel Wright, Zhisheng Xiao, Keyang Xu, Marc van Zee, Junlin Zhang, Wenlei Zhou and Konrad Zoln.

Contributors

Ola Aboubakar, Canfer Akbulut, Javier Lopez Alberca, Nina Anderson, Marco Andreetto, Lora Aroyo, Burcu Karagol Ayan, Ben Bariach, Sherry Ben, Dana Berman, Irina Blok, Pankil Botadra, Jenny Brennan, Karla Brown, Elie Bursztein, Viral Carpenter, Norman Casagrande, Ming-Wei Chang, Solomon Chang, Shamik Chaudhuri, Tony Chen, John Choi, Yu-Chuan Su, Dmitry Churbanau, Nathan Clement, Matan Cohen, Forrester Cole, Vincent Du, Praneet Dutta, Tom Eccles, Ndidi Elue, Ashley Feden, Shlomi Fruchter, Frankie Garcia, Roopal Garg, Ahmed Ghazy, Bryant Gipson, Dawid Górny, Yoni Halpern, Susan Hao, Amir Hertz, Ed Hirst, Tingbo Hou, Mohamed Ibrahim, Dirichi Ike-Njoku, Vlad Ionescu, William Isaac, Xuhui Jia, Gemma Jennings, Donovon Jenson, Kerry Jones, Yelin Kim, Suraj Kothawade, Jolanda Kumakaw, Dana Kurniawan, Dmitry Lagun, Tao Li, Maggie Li-Calis, Yuchi Liu, Kristian Lum, Chase Malik, John Mellor, Inbar Mosseri, Tom Murray, Aida Nematzadeh, Paul Nicholas, João Gabriel Oliveira, Michela Paganini, Roni Paiss, Alicia Parrish, Anne Peckham, Tobias Pfaff, Alex Pirozhenko, Ryan Poplin, Utsav Prabhu, Yuan Qi, Cyrus Rashtchian, Charvi Rastogi, Amit Raul, Ali Razavi , Susanna Ricco, Felix Riedel, Dirk Robinson, Pankaj Rohatgi, Bill Rosgen, Sarah Rumbley, Anthony Salgado, Florian Schroff, Candice Schumann, Tanmay Shah, Kaushik Shivakumar, Dennis Shtatnov, Zach Singer, Thibault Sottiaux, Brad Stone, Eric Tabellion, Shuai Tang, David Tao, Kurt Thomas, Andeep Tor, Aayush Upadhyay, Cristina Vasconcelos, Andrey Voynov, Amanda Walker, Miaosen Wang, Simon Wang, Stanley Wang, Qifei Wang, Yuxiao Wang, Olivia Wiles, Mete Yurtoglu, Andrew Xue, Ali Zand, Han Zhang, Catherine Zhao, Miao Zhou, Shengqi Zhu and Zhenkai Zhu

Advisors

Dawn Bloxwich, Mahyar Bordbar, Luis C. Cobo, Eli Collins, Tulsee Doshi, Anca Dragan, Douglas Eck, Nando de Freitas, Demis Hassabis, Tom Hume, Koray Kavukcuoglu, Helen King, Kathy Meier-Hellstern, Oriol Vinyals and Yori Zwols