In the video, Deva from the Imagen team showcases Imagen 2.0’s enhanced capabilities, including flexible aspect ratios, higher resolutions up to 2K for clearer and more detailed images, and intelligent aspect ratio inference based on prompts. These improvements enable the creation of tailored, high-quality images such as seamless 360-degree panoramas and detailed educational posters, significantly advancing the model’s usability and output quality.
In this video, Deva, a researcher on the Imagen team, introduces the new capabilities of Imagen 2.0 related to aspect ratios and resolutions. Unlike previous models that were limited to fixed portrait, landscape, or square formats, Imagen 2.0 allows users to generate images with variable widths and heights. This flexibility enables users to tailor images more precisely to their needs, whether for creative projects or practical applications.
Deva demonstrates the new features by creating a tall 3 by 1 poster illustrating the different layers of the ocean. This example highlights the importance of resolution, especially for printed materials where legibility is crucial. Previously, the API limited image resolution to a maximum of 1K, but now users can request up to 2K resolution. The higher resolution results in much crisper images, making small text clear and readable, which is ideal for educational purposes like classroom posters.
The video also showcases a comparison between lower and higher resolution images, emphasizing the significant improvement in clarity. For instance, text on a grain of rice is barely legible at the lower resolution but becomes clearly readable at the higher 2K resolution. This enhancement marks a substantial upgrade from earlier versions of the model, improving the usability of generated images in detailed and text-heavy contexts.
Another exciting feature of Imagen 2.0 is its ability to infer suitable aspect ratios based on the prompt. Deva experiments with generating a 360-degree panorama from the top of Half Dome. The model correctly produces an image with an approximate 2 by 1 aspect ratio, typical for panoramas. This intelligent aspect ratio selection simplifies the creation of specialized image formats without requiring manual input from users.
Finally, Deva tests the panorama by viewing it in a panorama viewer to check for seams. The model performs impressively, producing a seamless 360-degree image where the edges align perfectly. This seamless integration demonstrates the model’s advanced understanding of image continuity and consistency. Deva invites viewers to try and find the seam, confident that the model’s output is smooth and well-crafted.