Google unveils Imagen Editor & EditBench: AI-Powered text-based image editing tools with localized capabilities

Google has recently unveiled Imagen Editor, an innovative research diffusion-based model text-based image editing tool powered by AI. Building upon their existing image generation tool, Imagen, Google has taken a leap forward by introducing localized editing capabilities. Imagen Editor enables users to make precise adjustments to specific elements within an image, all through the power of textual prompts and sketching.

Imagen Editor allows users to modify existing photos by providing the image to be edited, a text prompt specifying the desired change, and a masked region indicating where the edit should take place. The edits performed by Imagen Editor are seamlessly integrated into the defined region, producing photorealistic and natural-looking results. Employing a technique known as inpainting, which involves image restoration, the tool resembles Adobe Photoshop's Content Aware Fill.

The researchers behind Imagen Editor implemented advanced encoders and incorporated an object detector module into the AI system. These enhancements effectively compensate for incomplete or inaccurate masks, ensuring optimal editing accuracy. Moreover, the introduction of EditBench, an evaluation tool for text-guided inpainting results, allows for a comprehensive analysis of edits on both human-made and AI-generated images. Parameters such as modified objects, their attributes, and suitability for the scene are considered when comparing the output.

While Imagen Editor showcases remarkable capabilities, Google remains committed to responsible AI development. Consequently, the company has made the decision not to release Imagen Editor to the general public. In an effort to establish ethical boundaries, Google proposes a safeguarding framework that sets strict limits on the deployment of such powerful tools. However, EditBench is available free of charge to support further research in the field. There is also anticipation among users that Imagen, the underlying model, will eventually be integrated into Gboard, offering an enhanced image generation experience for all.