Google I/O 2024: Text-to-Image AI Model Imagen 3 Unveiled, Gets Improved Image Generation Capabilities

Google made several new announcements at its annual developer-focused Google I/O 2024 event. Among many artificial intelligence (AI) focused announcements made during the keynote session, one was particularly surprising. The tech giant introduced the next generation of its text-to-image AI model, Imagen 3. The new AI model was introduced just months after the launch of its predecessor Imagen 2, which came out in December 2023 and was later upgraded last month. The company said the new model can generate detailed photorealistic images while closely following the prompt.

Imagen 3 was introduced by Douglas Eck, Senior Research Director at Google DeepMind. Unveiling it, he said, “Today, I’m so excited to introduce Imagen 3. It is our most capable image generation model yet. It understands prompts written the way people write. The more creative and detailed you are, the better. Plus, this is our best model yet for rendering text which has been a challenge for image generation models.”

The AI model’s ability to understand prompts is said to have been heavily improved, which now allows it to closely follow the prompt to capture small details and generate a faithful image. This also appears to be a common direction for most of the AI-related announcements during the event, as most of the AI models are now capable of better understanding prompts. Google added that Imagen 3 will be available in multiple versions where each model is optimised for a specific type of task that can range from generating quick sketches to creating high-resolution images.

To enable Imagen 3 to capture small details and specific instructions such as camera angles or compositions in long, complex prompts, Google has trained the AI model with images that contain detailed descriptions in its captions, allowing it to pick up on even smaller nuances. It can also generate a variety of textures and can render text-based images.

Focusing on safety, every image generated by Imagen 3 will contain its SynthID’s watermark labelling. It embeds a digital watermark directly into the pixels of the image, making it impossible to remove via cropping, sharing, or making any alterations to the image. The AI model is expected to arrive in a public preview in the coming months. Right now, Google is working on adding inpainting and outpainting editing options. Imagen 3 is currently available in private preview inside ImageFX for select creators. It will soon be made available for the tech giant’s enterprise customers.