LensErase from Microsoft Erases Watermarks and Objects from Images

Microsoft has unveiled the LensErase model, designed to remove watermarks, logos, text overlays, and objects along with their shadows from images. The tool operates by taking an input image and a selection mask, without requiring text prompts.

Key Features

The model is trained to recognize and remove various types of overlays: semi-transparent watermarks, repeating symbols, diagonal text, and corner markings. It can handle both fine details and large objects covering significant portions of the frame, including people, animals, and furniture.

• Simultaneous removal of multiple objects in different parts of the image
• Elimination of associated shadows and reflections extending beyond the selection boundary
• Text erasure while preserving the underlying surface and adjacent elements
• Editing existing posters by removing titles and logos

The TraceFusion technology is responsible for restoring the rest of the image by identifying the actual area of change and preserving the original content outside that boundary. This prevents unwanted processing of the entire frame when requesting the removal of a small element.

Model Training

For training, researchers assembled a dataset of over 101 million examples in the format 'original image, mask, clean result'. About 10 million of these examples contain watermarks, and separate datasets were also used for text, posters, and large removal areas.

A distinctive feature of this approach is that the model uses visible background details seen through semi-transparent overlays to reconstruct the image. The appearance of the removed object helps the algorithm determine the归属 of shadows and reflections.

Current Project Status

At present, the model's code is not available; authors have announced its imminent release. The repository lists limitations: exact restoration of lost details in occluded areas is not guaranteed, and preparing high-quality training data for ultra-high resolutions (including 4K) presents challenges.