LensErase: A Foundation Model for Erase Anything

1 Introduction

• Real-world erasing must handle targets from tiny regions to nearly half a scene, including text, watermarks, multiple objects, and associated effects.These cases require both precise local edits and reconstruction of large missing regions while maintaining scene consistency.Source: §1 · p. 2
• LensErase-100M provides 101.07 million paired examples spanning diverse categories and erase settings, with domain-specific pipelines for reliable source–target supervision.The dataset covers visual text, watermarks, object scales, resolutions, and multi-object cases.Source: §1 · p. 2
• LensErase adapts Lens’s generative prior to mask-based erasing, encoding region identity in source-image tokens without a separate mask-token sequence.Both branches start from Lens’s pretrained image branch, and the condition branch receives the intact source image.Source: §1 · p. 2
• Supervised fine-tuning, region-aware reinforcement-learning rewards, and TraceFusion refine removal quality and preserve content outside the effective edit region.TraceFusion estimates that region from the selection mask and source–output residuals, then selectively restores the source image outside it.Source: §1 · p. 2
• LensErase achieves substantial gains across six public benchmarks and LensErase-Bench under both reference-based objective metrics and reference-free VLM evaluation.The benchmark spans diverse erase scenarios, and the model is evaluated against 12 existing methods.Source: §1 · p. 3

2.1 Pre-Training Data: LensErase-100M

• LensErase-100M contains 101.07M image–mask–target triplets across General, Watermark, Large-Mask, Visual Text, and Poster data.Figure 2 presents distributions of mask-area ratios and mask counts, without values specified in the supplied text.Source: §2.1 · p. 3, Figure 2 · p. 3
• General Data reverses inserted-object edits to use the original real image as the clean target, while balancing sampled object categories and classes.Difference maps and SAM2 prompts produce the selection mask.Source: §2.1 · p. 3
• Specialized pipelines construct watermark triplets from synthesized overlays, visual-text triplets from screened teacher edits, and poster pairs by deleting selected layers.Poster masks include only visible pixels of deleted elements; watermark masks use the rendered overlay region.Source: §2.1 · pp. 4–5

2.2 Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) Data

• The 140K-example SFT set prioritizes high-resolution data and excludes synthetic Large-Mask examples, while RL uses 19K image–mask tasks with source images disjoint from SFT.RL data includes hard cases and targeted supplements; General Data requires no paired target, while Poster and Watermark use deterministic references and Visual Text uses verified weak references.Source: §2.2 · p. 5

2.3 Evaluation Dataset: LensErase-Bench

• LensErase-Bench contains 500 curated cases without paired ground-truth targets, balancing application scenarios, mask area, resolution, target count, and erase difficulty.A VLM evaluates outputs against the source image and selection mask.Source: §2.3 · p. 5
• The benchmark scores removal, inpainting, preservation, and cleanliness from 0–4, rescales them to 0–100, and reports a weighted overall score and HiFi Removal rate.A valid HiFi Removal case requires Removal = 4 and each other dimension to be at least 3 on the original scale.Source: §2.3 · pp. 5–6

3 Method

• LensErase takes a source image and binary mask, then removes selected objects and associated effects while preserving the scene and plausibly completing removed regions.Source: §3 · p. 6

3.1 Pretraining: Adapting Generation to Erasure

• Both Lens branches start from Lens's image branch, and Mask-Aware RoPE encodes masked status in source-token positions without extra mask tokens or embeddings.The model retains the selected region because its appearance and location help identify associated effects, while semi-transparent overlays retain information about the background.Source: §3.1 · pp. 6–7
• Pretraining uses region-weighted flow matching to emphasize masked regions, then supervised fine-tuning retains the same objective.The dilated mask is downsampled to the token grid, with λ = 5; loss normalization reduces sensitivity to mask size.Source: §3.1 · p. 7

3.2 RL: Contract-Guided Erasure

• A shared source-grounded erase contract specifies targets and effects to remove and nearby objects to preserve, guiding evaluation of stochastic candidates.The reward evaluates removal, reconstruction, preservation, cleanliness, and outside fidelity; removal-aware gating penalizes applicable removal failures so other criteria cannot fully compensate.Source: §3.2 · pp. 7–8

3.3 Post-Processing with TraceFusion

• TraceFusion estimates the effective edit region from source-output residuals, keeps generated content there, and restores the source elsewhere.Its estimated region can extend beyond the input mask to include effects such as shadows or reflections; residual detection uses thresholds and a supplementary gamma-transformed branch.Source: §3.3 · pp. 8–9

4 Experiments

• The experiments compare LensErase with 12 publicly released methods across seven benchmarks, using reference-based metrics on six public benchmarks and the LensErase-Bench evaluation metrics.The OmniPaint-Bench evaluation uses 300 captured cases and excludes its RORD-derived subset to avoid overlap with RORD-Val-343.Source: §4 · p. 9
• The mask-injection comparison evaluates Mask-Aware RoPE against Channel Concat and Token Concat at the pretraining stage.Both alternatives use pixel-unshuffled 16×16 mask patches and positional indices independent of mask values.Source: Table 4 · p. 10, §4 · p. 10
• LensErase achieves state-of-the-art performance across six public benchmarks.The reported main-table comparison shows the three baselines with the lowest mean rank per benchmark; full comparisons cover all 12 baselines.Source: §4 · p. 10
• LensErase achieves the best overall performance against 12 publicly released baselines on LensErase-Bench.The main table reports six representative baselines, with complete comparisons in the appendix.Source: §4 · p. 10, Table 2 · p. 10
• HiFi Removal rises from 65.80% after pretraining to 70.80% after SFT, 81.60% after RL, and 82.80% with TraceFusion.TraceFusion also increases mean PSNR from 28.70 to 30.82 dB and reduces LPIPS from 0.0816 to 0.0684.Source: §4 · p. 10, Table 3 · p. 10

A.1 Image Inpainting and Object Removal

• Classical inpainting uses boundary propagation or patch retrieval, while later methods add semantic completion, structural priors, and broader context.Diffusion-based removal methods use varied conditioning or guidance, while some approaches redirect attention or latent trajectories [48] [57] [61] [62] [63].Source: §A.1 · p. 17

A.2 Data and Models for Image Editing

• Earlier editing-data work ranges from manually annotated triplets to synthetic pairs and automated mining, while image-editing systems also study composition, customization, and spatial control [69] [70] [71] [72].The cited work includes human-in-the-loop refinement and context preservation during real-image editing [72].Source: §A.2 · p. 17
• LensErase uses a unified source–mask interface for natural objects, visual text, translucent watermarks, and layered designs, supported by domain-specific data and post-training.Its approach adapts a pretrained text-to-image prior through visual initialization and mask-aware conditioning, then refines removal and preservation.Source: §A.2 · p. 17

A.3 Object Effects and Background Preservation

• LensErase conditions on the intact source image to retain information about the selected target, associated effects, and content visible through translucent overlays; TraceFusion combines generated content with the original outside estimated edit support.Related methods address effects through counterfactual supervision, paired video frames [4], target-aware attention, or other mechanisms [74].Source: §A.3 · p. 17

A.4 Reward-Based Post-Training

• LensErase uses DiffusionNFT with removal-specific rewards assessing removal, reconstruction, semantic preservation, visual cleanliness, and fidelity outside the edit region.DiffusionNFT is one of several reward-optimization approaches, alongside policy-gradient and trajectory-backpropagation methods.Source: §A.4 · p. 18

B Data Construction and Filtering Details

• The data pipelines produce aligned input, output, and mask triplets; pretraining and SFT use such triplets, while RL uses curated image–mask tasks with references for reward computation.Figure 4 summarizes 101.07M pretraining examples, 140K SFT examples, and 19K RL tasks.Source: B Data Construction and Filtering… · p. 18

B.1.1 General Data: addition-to-removal counterfactuals

• General Data reverses object additions: an editing teacher adds an object to a real image, and the original image becomes the removal target.The instruction is used only by the teacher, and the aligned pair is represented as Iinput = Gadd(Ioutput, P), M = RecoverMask(Iinput, Ioutput).Source: §B.1.1 · p. 18
• Mask recovery combines low-frequency DCT changes, DINOv3 semantic differences, and SAM2 boundary refinement; the paired target can supervise removal of associated effects beyond the selection mask.Cross-signal agreement and outside-mask consistency are among the filters for unreliable pairs.Source: §B.1.1 · pp. 19–20

B.1.2 Alpha-supervised watermark synthesis

• Watermark pairs composite rendered overlays onto clean images, deriving the selection mask directly from the rendered alpha support rather than image differences.The synthesis includes localized and repeated marks, with randomized placement, scale, opacity, and appearance.Source: §B.1.2 · pp. 20–21

B.1.3 Generated large-foreground compositing

• Large-foreground pairs composite generated, segmented objects onto real backgrounds, varying visible area while retaining the pre-compositing image as the clean target.Placements target a sampled mask-area ratio of 20%–60%, with constraints limiting excessive occlusion.Source: §B.1.3 · pp. 21–22

B.1.4 OCR-verified partial visual text removal

• Visual Text Data selects OCR-eligible text groups and blends teacher reconstructions only into selected regions, preserving unselected text and surrounding source content.OCR checks residual text, while structural-change and gradient-discrepancy criteria screen removal and paste-back quality.Source: §B.1.4 · pp. 22–23

B.1.5 Layer-aware poster rendering

• Poster pairs reuse the same rendered layers and background, omitting selected layers for the target while keeping retained content fixed.Masks cover selected element bodies, although the paired images remove the full selected layers, including effects beyond the mask.Source: §B.1.5 · pp. 23–24

B.2 SFT Data Curation and Quality Screening

• SFT examples draw from General, Watermark, Visual Text, and Poster Data, with high-resolution candidates and professional posters included in curation.Triplets are screened for mask validity, complete intended removal, preservation of unrelated content, and plausible reconstruction without conspicuous artifacts.Source: §B.2 · p. 24

B.3 RL Data Pipeline

• The 19K RL tasks comprise 13K General examples and 2K each for Poster, Watermark, and Visual Text; domains are sampled at 6:2:1:1 during RL, independently of dataset sizes.General Data includes 8K difficulty-filtered and 5K supplementary tasks; reward specifications do not change the model’s source-image–mask interface.Source: §B.3 · pp. 24, 25, 27
• General Data supplements difficulty-filtered tasks with failure-guided collection and attributable object effects, while masks select objects rather than exhaustively marking every pixel that should change.Object bodies are segmented from verified boxes, and effects are not automatically added to masks; difficulty is assessed after mask construction.Source: §B.3 · p. 25
• 6,991 Hard and 1,009 Extreme tasks form the 8K difficulty-filtered subset; five weighted difficulty ratings guide selection, not the training reward.The ratings cover removal scope and effects, reconstruction, occlusion, preservation risk, and unwanted-entity risk.Source: §B.3 · pp. 25–26
• Poster Data contains 2,000 unreduced-resolution designer examples, each removing one visible non-background layer, with a clean reference rendered from the same layer stack.The selected layer may be behind foreground layers; masks derive from final RGB differences, excluding unchanged or occluded pixels.Source: §B.3 · p. 26
• Watermark Data uses 2,000 distinct high-resolution backgrounds with synthesized overlays and exact clean references, balancing six marginal attributes including layout, area, opacity, color, and mask style.Half the tasks use precise masks and half use coarse local selections that cover all actual changes.Source: §B.3 · p. 26
• Visual Text Data contains 2,000 OCR-verified partial-removal tasks with weak references; target-centered crops enlarge small text while keeping source, mask, and reference aligned.Because the hidden text-free appearance is not directly observed, reference-guided rewards do not require every plausible completion to match the teacher reconstruction.Source: §B.3 · pp. 26–27

C LensErase-Bench: Construction and Evaluation

• LensErase-Bench contains 500 cases across General, Poster, Watermark, and Visual Text, covering varied mask scales, target counts, resolutions, and removal difficulty.Its domain counts are 200 General and 100 each for the other three domains.Source: C LensErase-Bench: Construction and… · p. 27

C.1 Selection and Coverage

• The benchmark balances resolution ranges across domains and varies mask area and domain-specific attributes; 59.2% of cases have one target and 40.8% have multiple targets.General cases split evenly by effects and target count; Poster, Watermark, and Visual Text subsets cover their respective layer, overlay, language, layout, and background attributes.Source: §C.1 · pp. 27–28, Table 5 · p. 27, Table 6 · p. 28

C.2 Evaluation Protocol

• Each case receives 0–4 ratings for Removal, Inpainting, Preservation, and Cleanliness from aligned views without a clean reference; method scores average dimensions equally across valid cases.Dimension ratings are mapped to 0–100, while HiFi Removal reports the fraction meeting all removal and editing-quality requirements.Source: §C.2 · pp. 28–29

D.1 Visual Initialization and Mask Conditioning

• Both LensErase branches use visual inputs initialized from Lens’s image branch, and Mask-Aware RoPE encodes masked versus unmasked source tokens without a separate mask embedding.The binary mask is pooled to the source-token grid and selects modality coordinates; RoPE is applied to queries and keys in every block.Source: §D.1 · p. 29

D.2 Region-Weighted Objective Details

• The flow-matching objective uses the original selection mask for conditioning but a separately dilated mask to define loss support.The loss-support mask uses a 31 × 31 pixel-space dilation before pooling and thresholding at the token grid.Source: §D.2 · p. 30

D.3 Two-Stage Pretraining Schedule

• Stage 1 trains the condition stream while freezing image-specific parameters; Stage 2 unfreezes the image backbone and jointly optimizes both streams.The schedule is intended to adapt the visual condition stream without immediately disrupting the pretrained image-generation prior.Source: §D.3 · p. 30, Figure 5 · p. 30

D.4 Initialization and Training Ablations

• 65.80% HiFi Removal for Mask-Aware RoPE exceeds Channel Concat (58.80%) and Token Concat (51.40%).Table 7 compares conditioning and initialization designs; its shared reference uses Mask-Aware RoPE, image-branch initialization, and two-stage training.Source: §D.4 · p. 30, Table 7 · p. 31
• 65.80% HiFi Removal for freeze–unfreeze exceeds No freeze (55.80%) and Always freeze (58.60%).The comparison supports adapting the visual condition stream first, then jointly adapting both streams.Source: §D.4 · pp. 30–31, Table 7 · p. 31

E.1 Supervised Fine-Tuning

• SFT uses LoRA while keeping VAE and backbone weights fixed, retaining the image–mask interface and region-weighted flow-matching objective.The supplied implementation details specify rank-64 LoRA, 5,000 optimization steps, and a learning rate of 2 × 10^-5.Source: §E.1 · p. 32

E.2 Contract-Guided Reward Construction

• For General Data, reward evaluation first fixes a candidate-independent contract of targets, attributable effects, and nearby content to preserve, then scores each candidate against it.Candidates share the same contract and detail-view geometry for a given image–mask condition.Source: §E.2 · p. 32, Figure 6 · p. 32
• Normalized VLM and fidelity subscores are aggregated over applicable criteria, with weights renormalized when criteria or components are inapplicable, before removal-aware gating.A zero-valued applicable Removal subcriterion contributes a factor of 0.5, so multiple failures compound without a hard rejection based on one zero.Source: §E.2 · pp. 32–34
• Outside Fidelity uses source–candidate PSNR over unchanged regions, excluding selection, declared effects, and qualifying residual components subject to valid-area thresholds.The supplemental residual exclusion is discarded if it covers more than 45% of the valid canvas; the metric requires at least 1,024 pixels and 8% of that canvas.Source: §E.2 · p. 33
• Poster, Watermark, and Visual Text rewards combine reference-based measures with VLM judgments, while Visual Text allows plausible reconstructions to differ from its weak reference.Their component weights are Removal 0.20, Reconstruction 0.20, Preservation 0.15, Outside Fidelity 0.25, and Cleanliness 0.20.Source: §E.2 · p. 34, §E.2.1 · p. 34

E.2.1 Reward Prompt Excerpts

• General Data assessment treats selected objects and clearly attributable effects as removal targets, while jointly checking all targets and independent non-target objects.A clear failure on one target cannot be offset by successful removal of other targets.Source: §E.2.1 · p. 34

E.3 DiffusionNFT Implementation

• DiffusionNFT trains on condition groups with sufficient valid-reward variation, using fixed SFT references, adapter updates, and 100 successful policy updates.Groups require at least 12 of 16 valid candidates and nonzero reward spread; collections with σB < 10^-3 are skipped.Source: §E.3 · p. 35

E.4 Inference and Sampling

• Inference uses 20 deterministic Euler steps from t = 1 to t = 0, with source encoding and mask held fixed.The VAE decodes only the terminal latent, and the update rule has no learned parameters of its own.Source: §E.4 · pp. 35–36

E.5 TraceFusion Parameters

• TraceFusion estimates the effective edit region from a smoothed source-output residual, then uses that support to guide blending.Its detector uses resolution-scaled thresholds and morphology; the same parameters are used for all reported experiments.Source: §E.5 · pp. 36–37

F.1 Evaluation Protocol

• Evaluation gives all methods the same source images and masks, scores restored-size outputs, and uses full-image resizing for CMMD.OmniPaint-Bench evaluation corrects 32 misaligned masks while leaving the images and targets unchanged.Source: §F.1 · p. 38, §F.1.1 · p. 38

F.1.1 OmniPaint-Bench Mask Corrections

• The OmniPaint-Bench protocol rotates 29 masks and flips three to align their orientations with input images; the remaining 268 masks are unchanged.All methods receive the same corrected masks, and the transformations preserve mask pixels and foreground area.Source: §F.1.1 · p. 38

F.2 Baseline Configurations

• Baseline configurations use released model checkpoints or specified variants, with repository and weight links provided for reproducibility.The appendix lists configurations for object-removal baselines including ObjectClear, SmartEraser, OmniPaint, OmniEraser, and OSOR.Source: §F.2 · p. 38

F.3 Results on Public Benchmarks

• Tables 10–12 expand the OBER-Test, RORD-Val-343, and RemovalBench comparisons to all 12 baselines across eight metrics.The tables mark the best, second-best, and third-best values among 13 methods.Source: Table 10 · p. 39, Table 11 · p. 39, §F.3 · p. 39
• Tables 13–15 provide the corresponding comparisons for captured-subset OmniPaint-Bench, AnimeEraseBench, and TextEraseBench.The OmniPaint-Bench table excludes RORD-derived cases; the expanded tables use the same protocol and eight metrics.Source: §F.3 · p. 39, Table 13 · p. 40, Table 14 · p. 40

F.4 Results on LensErase-Bench

• PANDORA is evaluated on 499 LensErase-Bench cases because one output was rejected by the safety filter; all other methods use 500.HiFi Removal counts cases with complete removal and Inpainting, Preservation, and Cleanliness scores of at least 3.Source: Table 16 · p. 41, §F.4 · p. 41

Reward ablations

• The five ablation variants omit Removal, Reconstruction, Preservation, Outside Fidelity, or Cleanliness, with remaining weights renormalized.The variant without Removal also disables the removal-aware gate; all variants start from the same SFT checkpoint.Source: G Reward Component Ablations · p. 41
• Removing the Removal component and its gate drops Removal from 94.95 to 43.35 and HiFi Removal from 81.60% to 38.80%.This joint removal produces the largest degradation among the tested variants.Source: G Reward Component Ablations · p. 41, Table 17 · p. 42
• The full reward achieves the best HiFi Removal; removing Preservation yields 74.60%, while removing Reconstruction or Cleanliness yields 78.00% or 78.40%.Removing Outside Fidelity produces the largest PSNR drop, from 28.70 to 26.69 dB.Source: Table 17 · p. 42, G Reward Component Ablations · p. 42

H Limitations and Conclusion

• Existing general-purpose editing models are less reliable for ultra-high-resolution editing, particularly 4K, constraining the scale and consistency of training data.The authors suggest more reliable high-resolution data construction and more challenging erase scenarios as future work.Source: H Limitations and Conclusion · p. 42

I.1 Comparisons with Mask-based GPT-Image-2.5-Sunburst [32] and Flux2-Erase [33]

• Figures 7–10 are labeled as qualitative comparisons with GPT-Image 2.5 and Flux2-Erase; their captions identify the compared systems but do not state which system performs better.Source: Figure 7 · p. 43, Figure 8 · p. 44, Figure 9 · p. 45

I.2 Comparisons with Object-Removal Baselines

• Appendix I.2 shows qualitative comparisons with object-removal baselines on general images, posters, watermarks, and visual text.The supplied figure captions specify these example types without reporting comparative outcomes.Source: Figure 11 · p. 47, Figure 12 · p. 48, Figure 13 · p. 49

• Figures 18–21 show LensErase results using hand-drawn masks and SAM 2 segmentations.Source: J Removal with Hand-Drawn and SAM Masks · p. 54
• Examples span text, transparent objects, multiple targets, and associated shadows or reflections.Source: J Removal with Hand-Drawn and SAM Masks · p. 54
• Each row presents the input image, mask, and LensErase output.Source: J Removal with Hand-Drawn and SAM Masks · p. 54