Google’s AI-based Imagen Editor Understands Your Descriptions
Google continues pushing the envelope in artificial intelligence (AI) and machine learning with their latest addition, Imagen Editor. This innovative text-based image generation tool allows local edits on photos by leveraging both textual prompts and user sketching. Although it won’t be released to the public due to responsible AI concerns, it represents a significant advance in AI-assisted photo editing.
Text-Guided Image Editing Revolution
In recent years, Google’s focus on AI has led them towards image creation applications, including tools like Midjourney and DALL-E 2. While their Imagen utility can create images from scratch using only textual prompts, it lacks the ability to edit specific image elements that users may not be satisfied with.
The company has recently shared research for Imagen Editor and EditBench, which are still in beta but have shown promising results in guiding edits with text prompts. By combining an existing photo with a text prompt describing the desired change, users can apply precise edits limited to a defined region of the image. The resulting alterations are photorealistic and natural.
How Imagen Editor Works
Technically termed inpainting, the method used by Google’s new Inpainting tool can be likened to a hybrid of Google AI and Adobe Photoshop’s Content Aware Fill. This process involves restoring or modifying an image based on the user’s input.
To improve Imagen Editor’s capabilities, researchers have developed new encoders and incorporated an object detector module within the AI. This addition compensates for any incomplete or inaccurate masks provided by users when selecting the region they want to edit.
Introducing EditBench
As part of their research, Google has also introduced EditBench, a tool used for evaluating the results of text-guided inpainting. The benchmark, which is based on a 240-image dataset, assesses human-made and AI-generated images regarding modified objects, attributes (shape, size, number), and suitability for the scene.
Google’s Imagen Editor outperformed alternatives like DALL-E 2 and StableDiffusion in all categories tested by EditBench. The company found that object masking improves image-text alignment, resulting in higher-quality image editing.
Responsible AI Concerns Restrict Public Release
Despite its promising capabilities, Google has decided not to release Imagen Editor to the public due to unspecified concerns regarding responsible AI use. The company aims to establish safeguards and hard limits before granting widespread access to AI-powered tools like Imagen Editor.
However, EditBench will be available in its entirety for free as a contribution to further AI research. Meanwhile, we can hope that the base model of Imagen might be integrated into Gboard in the near future.
Practical Applications Abound
If released to the public, Imagen Editor offers several practical applications and user-friendly photo editing experiences. Users who lack expertise in advanced photo editing software would be able to make complex edits quickly and efficiently with this tool.
Aside from addressing various visual elements in photographs or virtual art projects, it would prove useful for simple tasks such as removing unwanted elements from backgrounds or changing specific colors.