LensErase: A Foundation Model for Erase Anything

LensErase. A Foundation Model for Erase Anything. By Yunuo Chen and thirteen others, from Shanghai Jiao Tong University, Nankai University, and four other institutions. Published on the thirteenth of October. Its code repository is public but slight: two source files.

Erasing something from an image covers far more ground than removing one object. Real-world erase requests involve visual text, watermarks, the visual effects associated with an object, several objects at once, and every range of object scale, aspect ratio and image resolution. A model meant to handle all of that needs coverage that a narrow erase tool does not have.

LensErase is offered as a general-purpose, mask-based foundation model for Erase Anything, designed to span that whole spectrum of scenarios. It is built in three moves: a very large dataset, pretraining from an existing generative model, and a refinement stage.

The dataset is LensErase-100M, containing one hundred and one point zero seven million paired examples of source, mask and target, spread across diverse erase scenarios. Each example shows the model what to remove and what the result should look like, and the scale is what lets one model see all of those scenarios. The model is pretrained from the text-to-image model Lens, adapting its strong generative prior to the job of mask-based erasing.

To bring in the source image and the erase mask efficiently, the authors propose Mask-Aware RoPE, a three-dimensional rotary positional encoding that injects mask information directly into the source-image tokens, with no separate sequence of mask tokens. The mask rides along with the image tokens rather than adding a second stream for the model to attend to. The model is then refined with supervised fine-tuning and with reinforcement learning, using rewards that are contract-guided and region-aware. At inference, a step called TraceFusion preserves high-fidelity content outside the region actually being edited.

Evaluation spans six public benchmarks plus a new one, LensErase-Bench. Across them, LensErase achieves substantial improvements, both on objective metrics that compare against a reference and on reference-free evaluation by a vision-language model. Using both kinds of measure matters for erasing, where a good result has to look natural and also stay faithful to the original image outside the mask.

The authors hope LensErase can serve as a foundational model for erase-related applications and systems. Breadth is the organising principle throughout: a dataset sized to the variety of the task, a prior borrowed from a strong generator, and safeguards that keep untouched pixels untouched.