SFT samples
Higher-resolution pairs screened by GPT-5.6 Sol for mask validity, edit correctness, reconstruction, and preservation.
LensErase
A Foundation Model for Erase Anything
An image. A mask. A clean result.
Remove objects, text, watermarks, and their associated effects.
1 Shanghai Jiao Tong University2 Nankai University3 USTC4 University of Sydney5 Peking University6 Microsoft
OriginalSelectLensErase
LENSERASE IN ACTION
Erasure across scenes and designs
From layered posters to reflections and multiple objects. Watch each edit, or click a case to inspect the original, mask, and result.
THE DATA BEHIND LENSERASE
LensErase-100M
LensErase-100M provides aligned input–mask–target triplets across five categories.
101.07M
LensErase-100M
Paired pretraining examples
General77.6M
Watermark9.99M
Large mask4.02M
Visual text1.44M
Poster8.02M
RL samples
Curated image–mask samples across general images, posters, watermarks, and visual text, with domain-specific reward supervision.
EVALUATING THE INTENDED EDIT
LensErase-Bench
Removal can have more than one valid completion. Our 500-case benchmark evaluates outputs from the source image and mask, without requiring a paired ground-truth target.
200 General100 Poster100 Watermark100 Visual text
GPT-5.6 Sol judges four complementary dimensions.
Removal
Complete removal of the targets and associated effects.
40% of OverallInpainting
Plausible completion consistent with the surrounding context.
20% of OverallPreservation
Retention of non-target objects, structures, and content.
20% of OverallCleanliness
Absence of visible artifacts and unrelated additions.
20% of OverallHiFi Removal
A successful case achieves Removal = 4 and all other dimensions ≥ 3 on the original 0–4 scale.
QUALITATIVE COMPARISONS
Comparisons with commercial models
Comparisons with GPT-Image2.5-Sunburst (mask-based editing) and Flux2-Erase on general images, posters, watermarks, and visual text. Click an image to inspect the outputs at a larger size.
QUALITATIVE COMPARISONS
Comparisons with open-source models
Results alongside OSOR, OmniPaint, ObjectClear, GeoRemover, and RORem-mixed. Choose a single baseline for larger side-by-side views.
Image Credits
Image sources are linked in the previews.
PhotographyPexels · Wikimedia Commons
DatasetsCrello · EntitySeg · UHRSD · HRS10K
Wikimedia Commons credits
Synthetic watermarks were added to these source images; model-generated removal outputs are also shown.
Selection overlays and outputs are edits. These examples do not imply endorsement by the people depicted.