Microsoft researchers introduced LensErase, a foundation model for removing selected content from images while preserving surrounding areas. The paper was published on October 13 and submitted to Hugging Face on October 15. The model is designed for objects, visible text, watermarks and their visual effects, as well as removing multiple objects across different sizes, aspect ratios and image resolutions.

The researchers built LensErase-100M, a dataset of 101.07 million paired examples, each containing a source image, an erase mask and a target image. They pretrained LensErase from the text-to-image model Lens, adapting its generative prior for mask-based erasing. A method called Mask-Aware RoPE injects mask information directly into source-image tokens using 3D rotary positional encoding, without adding a separate sequence of mask tokens.

They then refined the model with supervised fine-tuning and reinforcement learning using contract-guided, region-aware rewards. During inference, TraceFusion is applied to preserve high-fidelity content outside the effective edit region. Across six public benchmarks and the researchers’ LensErase-Bench, the model showed substantial improvements on reference-based objective metrics and reference-free VLM evaluation; the abstract does not give specific scores. The authors said LensErase is intended to support erase-related applications and systems.

Microsoft researchers introduced LensErase, a foundation model for removing selected content from images while preserving surrounding areas. The paper was published on October 13 and submitted to Hugging Face on October 15. The model is designed for objects, visible text, watermarks and their visual effects, as well as removing multiple objects across different sizes, aspect ratios and image resolutions.

The researchers built LensErase-100M, a dataset of 101.07 million paired examples, each containing a source image, an erase mask and a target image. They pretrained LensErase from the text-to-image model Lens, adapting its generative prior for mask-based erasing. A method called Mask-Aware RoPE injects mask information directly into source-image tokens using 3D rotary positional encoding, without adding a separate sequence of mask tokens.

They then refined the model with supervised fine-tuning and reinforcement learning using contract-guided, region-aware rewards. During inference, TraceFusion is applied to preserve high-fidelity content outside the effective edit region. Across six public benchmarks and the researchers’ LensErase-Bench, the model showed substantial improvements on reference-based objective metrics and reference-free VLM evaluation; the abstract does not give specific scores. The authors said LensErase is intended to support erase-related applications and systems.