acceptodds
Under review as a conference paper at ICLR 2027

LensErase: A Foundation Model for Erase Anything

Abstract

We introduce LensErase, a general-purpose mask-based Erase Anything foundation model designed to cover a broad spectrum of real-world erase scenarios. LensErase supports diverse objects, visual text, watermarks, associated visual effects, multiple-object removal, and a wide range of object scales, aspect ratios, and image resolutions. First, to enable such broad coverage, we construct LensErase-100M, a large-scale dataset containing 101.07 million paired source–mask–target examples spanning diverse erase scenarios. Second, we pretrain LensErase from the text-to-image model Lens, adapting its strong generative prior to mask-based erasing. To efficiently incorporate the source image and erase mask, we propose Mask-Aware RoPE, a 3D rotary positional encoding that directly injects mask information into the source-image tokens without introducing a separate mask token sequence. Finally, we refine the model through supervised fine-tuning and reinforcement learning with contract-guided, region-aware rewards, and apply TraceFusion at inference to preserve high-fidelity content outside the effective edit region. Across six public benchmarks and the proposed LensErase-Bench, LensErase achieves substantial improvements under both reference-based objective metrics and reference-free VLM evaluation. We hope LensErase can serve as a foundational model for erase-related applications and systems.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
object erase, object remove

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_2eszOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LensErase, a model for removing objects and other content from images. It also constructs LensErase-100M, a dataset containing 101.07M samples covering objects, visual text, watermarks, and posters. The method is based on Lens and uses Mask-Aware RoPE to include mask information. Meanwhile, SFT, RL, and TraceFusion are used to improve the removal results and preserve unrelated content. Experiments on six public benchmarks and LensErase-Bench show that the method outperforms multiple baseline methods.

Critical strengths
  1. The dataset covers different removal tasks. The strategy of adding objects to real images and using the original images as clean targets is reasonable.
  2. The experiments cover six public benchmarks and 12 baseline methods. The results are strong. The ablation studies in Table 3 and Table 4 are also useful.
Critical weaknesses
  1. The paper uses 101.07M training samples and a large amount of computing resources. However, it does not report results with smaller training sets. Therefore, it is unclear how much of the improvement comes from the proposed method rather than the data scale.
  2. The RL training uses GPT-5.6 Sol to provide reward scores. Meanwhile, LensErase-Bench also uses the same model for evaluation. I am concerning whether there will be hacking with same reward model and judge model.
  3. In Appendix G, the Removal reward and removal-aware gating are removed together. Therefore, the independent contribution of the gating mechanism in Eq. (6) remains unclear.
Questions for authors

Please see the weaknesses above. I am willing to keep a positive rating if these concerns are properly addressed.

5 comments

Reviewer 2esz

We sincerely thank Reviewer 2esz for the thoughtful assessment and constructive suggestions. Below, we address each point with a controlled reduced-data pretraining study (500K vs. 101.07M), multi-judge evaluations alongside a 2,000-judgment human study, and an isolated ablation of the removal-aware gate.

Please feel free to let us know if any further discussions or experiments would be helpful.



1. Contribution of pretraining data scale

📝 Question. How much of the improvement comes from the method and how much comes from the 101.07M pretraining examples?

💡 Response. Due to the high training cost (two weeks on 96 A100 GPUs) and the limited rebuttal timeframe, we cannot run additional data-scale sweeps from scratch. Instead, we report a controlled reduced-data experiment from our earlier development, trained on 500K sampled pretraining triplets (~0.5% of LensErase-100M). All architecture designs, SFT data (140K), and RL configurations (19K) were kept identical.

This comparison highlights two clear findings:

  1. Strong method contribution at small data scale. Even with 200x less pretraining data, every stage already surpasses all existing baselines in Table 2 (where the strongest baseline, OmniPaint, scores 73.59 Overall and 42.80% HiFi Removal). Our 500K pipeline reaches 89.92 Overall and 71.20% HiFi Removal. Moreover, RL boosts HiFi Removal by +14.20 pp on the 500K model, confirming the effectiveness of our method.
  2. Substantial benefit from scaling to 101.07M. Scaling pretraining from 500K to 101.07M improves both Overall Score and HiFi Removal at every stage shown in the tables below, raising final Overall Score from 89.92 to 95.22 (+5.30) and HiFi Removal from 71.20% to 82.80% (+11.60 pp). Thus, our method design provides a strong foundation even at modest scale, while LensErase-100M unlocks large additional gains.

Pretraining on 101.07M triplets

StageRemovalInpaintingPreservationCleanlinessOverallHiFi Removal (%)
Pretraining82.3584.6591.0592.1586.5165.80
+ SFT88.6587.1592.0593.0589.9170.80
+ RL (Raw)94.9592.6593.3596.1594.4181.60
+ TraceFusion95.1592.4096.3597.0595.2282.80

Pretraining on 500K triplets

StageRemovalInpaintingPreservationCleanlinessOverallHiFi Removal (%)
Pretraining79.5075.5583.1588.8581.3151.00
+ SFT80.0577.2085.6087.9582.1756.60
+ RL (Raw)91.0086.8084.3094.6089.5470.80
+ TraceFusion90.6087.2586.3094.8589.9271.20

2. Potential reward overoptimization

📝 Question. Can RL optimize preferences specific to GPT-5.6 Sol and thereby increase the LensErase-Bench score?

💡 Response. The RL reward combines VLM judgments with pixel fidelity terms, while LensErase‑Bench scores the full output image using a separate rubric. To verify that RL improves genuine removal quality without overfitting to GPT-5.6 Sol preferences, we conducted three independent checks outside the training judge:

1. SFT-to-RL gains across unseen VLM judges.

We re-evaluated the exact same SFT and RL (Raw) outputs (before TraceFusion) across all 500 LensErase-Bench cases using three unseen judges: GPT-6 Astra, GPT-6.1 Sol, and cross-family Gemini 3.8 Flash. All three independent judges report consistent gains in Overall Score and HiFi Removal, closely matching the +10.80 pp under GPT-5.6 Sol. These steady improvements confirm that RL genuinely resolves incomplete removals and local artifacts, without relying on judge-specific biases.

JudgeSFT OverallRL (Raw) OverallΔ OverallSFT HiFi Removal (%)RL (Raw) HiFi Removal (%)Δ HiFi Removal (pp)
GPT-5.6 Sol (paper)89.9194.41+4.5070.8081.60+10.80
GPT-6 Astra91.4595.09+3.6470.2081.00+10.80
GPT-6.1 Sol92.0395.87+3.8472.8083.00+10.20
Gemini 3.8 Flash92.3196.69+4.3879.4090.00+10.60

2. Final method rankings across unseen VLM judges.

Using the exact same model outputs and scoring rubric as in Table 2, we re-evaluated LensErase and the six representative baselines across all 500 cases of LensErase-Bench with three independent VLM judges, including two newer models (GPT-6 Astra, GPT-6.1 Sol) and a cross-family judge (Gemini 3.8 Flash).

Across all four judges, LensErase consistently ranks 1st in both Overall Score and HiFi Removal ratio. Competing baselines also show only minor rank swaps among closely scored models, confirming the robustness of our evaluation rubric across VLM families.

Method · OverallGPT-5.6 Sol ScoreGPT-5.6 Sol RankGPT-6 Astra ScoreGPT-6 Astra RankGPT-6.1 Sol ScoreGPT-6.1 Sol RankGemini 3.8 Flash ScoreGemini 3.8 Flash Rank
LensErase (w/ TraceFusion)95.22195.29195.61196.871
OmniPaint73.59273.91475.04477.465
OSOR71.45475.37276.66280.883
ObjectClear71.67374.06376.26381.182
RORem-mixed69.75572.63574.35579.964
GeoRemover62.70660.39761.92767.747
OmniEraser58.89763.97665.93669.726
Method · HiFi Removal (%)GPT-5.6 Sol ScoreGPT-5.6 Sol RankGPT-6 Astra ScoreGPT-6 Astra RankGPT-6.1 Sol ScoreGPT-6.1 Sol RankGemini 3.8 Flash ScoreGemini 3.8 Flash Rank
LensErase (w/ TraceFusion)82.80181.00183.60190.201
OmniPaint42.80235.20238.20248.804
OSOR32.60333.20336.20354.402
ObjectClear28.40428.60434.00449.803
RORem-mixed27.40523.20526.00545.805
GeoRemover25.60620.80625.40637.406
OmniEraser22.60719.80723.60733.807

3. Blind human pairwise evaluation. We conducted a blind human pairwise study comparing LensErase with the six baselines in Table 2. Twenty participants each completed 100 comparisons (2,000 judgments in total). Given the marked input image and two anonymized outputs in randomized A/B order, participants chose the result with cleaner target and effect removal, more natural inpainting, and better background preservation (ties allowed).

Fitting a Bradley-Terry model [1] (ties as half-wins, centered at 1500 Elo), LensErase ranks 1st at 1770.5 Elo, leading the strongest baseline ObjectClear by 249.0 Elo with 465 wins, 47 ties, and 59 losses. The human ranking also aligns closely with the GPT-5.6 Sol rankings on LensErase-Bench.

[1] Bradley, Ralph Allan, and Milton E. Terry. "Rank analysis of incomplete block designs: I. the method of paired comparisons." Biometrika 39.3/4 (1952): 324-345.

MethodElo ↑Human RankWinsTiesLossesVLM Overall Score RankVLM HiFi Removal Rank
LensErase1770.51465475911
ObjectClear1521.522903724434
OmniPaint1507.432793425822
OSOR1483.642564127543
RORem-mixed1462.552393829555
GeoRemover1410.662022234766
OmniEraser1343.971433339677


Summary. Taken together, these three additional evaluations support improved removal quality of RL and show no signs of reward overoptimization specific to GPT-5.6 Sol.

3. Independent effect of the removal-aware gate

📝 Question. Does the ablation isolate the effect of the gate in Eq. (6)?

💡 Response. In Table 17 of the submission, w/o Removal and gate is a joint ablation. In Eq. (6), each zero-scored Removal subcriterion halves the aggregate RL reward. This gate is designed to prevent strong scores on other reward components (such as Preservation or Cleanliness) from offsetting clear removal failures. To isolate this gate, we trained an additional variant (w/o gate) that sets the gate multiplier to 1 while keeping all five reward components and weights unchanged, including Removal.

Comparing w/o Removal and gate , w/o gate , and Full reward separates the two effects. The Removal term provides the main gain (+38.00 pp), while the gate adds another +4.80 pp in HiFi Removal, +3.90 points in Removal, and +0.51 dB in mean PSNR across the six public benchmarks. Without the gate, RL can offset incomplete removals with higher Preservation and Cleanliness scores, which lowers the HiFi Removal rate.

Model variantRemovalInpaintingPreservationCleanlinessOverallHiFi Removal (%)Public PSNR ↑Public LPIPS ↓
Full reward (Raw)94.9592.6593.3596.1594.4181.6028.700.0816
w/o gate91.0589.1097.1597.4093.1576.8028.190.0837
w/o Removal and gate43.3546.1098.7098.1065.9238.8021.850.1446
w/o Reconstruction90.8591.4096.1095.7092.9878.0028.320.0791
w/o Preservation94.1591.5089.2096.0093.0074.6028.010.0817
w/o Outside Fidelity92.9091.7094.4096.4093.6678.0026.690.0888
w/o Cleanliness90.9090.9596.2595.9592.9978.4028.480.0803


We have incorporated these results and clarifications into the revised manuscript.

Reviewer_15ufOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces LensErase, a image-erasing model capable of removing objects, text, watermarks, and their associated visual effects. The authors constructed a dataset, LensErase-100M, comprising 101 million training samples, and proposed Mask-Aware RoPE to enable efficient mask-conditioned control. The model is further refined through SFT, reinforcement learning, and a post-processing technique called TraceFusion. Experimental results across six public benchmarks as well as on LensErase-Bench, a new benchmark introduced in this work demonstrate that the model achieves superior performance, outperforming existing methods.

Critical strengths
  1. LensErase-100M provides training dataset that encompasses a wide range of erasure scenarios.
  2. By integrating Mask-Aware RoPE, reinforcement learning fine-tuning, and TraceFusion technology, the framework establishes a practical and highly efficient image-erasure pipeline.
  3. This method demonstrates significant improvements over existing approaches across six public benchmarks, with its effectiveness further validated through ablation studies.
Critical weaknesses
  1. This paper reports results based on automated metrics but it does not include human user studies to verify the quality and visual realism.
  2. This post-processing method relies on heuristic thresholds, lacks a comprehensive sensitivity analysis, and does not evaluate potential failures or their associated impacts.
  3. In RL, using the same proprietary VLM as both the reward model and the evaluator for LensErase-Bench leads to a significant evaluation loop. Although RL yields only marginal improvements on public benchmarks (LPIPS: 0.0826 -> 0.0816), it boosts the model's HiFi Removal score on LensErase-Bench from 70.80% to 81.60%.
Questions for authors

see critical weaknesses

4 comments

We sincerely thank Reviewer 15uf for the thoughtful assessment and constructive suggestions. Below, we address each point with a 2,000-judgment human pairwise study, a hyperparameter and failure analysis of TraceFusion, and cross-benchmark as well as multi-judge evaluations for RL.

Please feel free to let us know if any further discussions or experiments would be helpful.



1. Human assessment of quality and realism

📝 Question. Can human judgments validate the visual quality and realism measured by the automated evaluation?

💡 Response. Yes. We conducted a blind human preference study comparing LensErase with the six representative baselines from Table 2. Twenty participants each completed 100 pairwise comparisons, yielding 2,000 judgments in total. In each trial, participants viewed the source image with the target region marked alongside two anonymized outputs in randomized A/B order. They selected the output that achieved cleaner removal of the target and its associated effects, more realistic inpainting, and better preservation of unrelated content (with ties allowed).

We fit a Bradley-Terry model [1] to all comparisons, counting ties as half-wins and centering ratings at 1500 Elo. LensErase ranks 1st at 1770.5 Elo, outperforming the strongest baseline (ObjectClear) by 249.0 Elo and winning 81.4% of its matchups outright (465 wins, 47 ties, 59 losses). Moreover, the human ranking closely matches both VLM rankings on LensErase-Bench: Ranks 1 and 5 to 7 are identical, while the same three closely performing baselines occupy Ranks 2 to 4. This directly validates the superior performance of LensErase and the reliability of our VLM evaluation rubric.

[1] Bradley, Ralph Allan, and Milton E. Terry. "Rank analysis of incomplete block designs: I. the method of paired comparisons." Biometrika 39.3/4 (1952): 324-345.

MethodElo ↑Human RankWinsTiesLossesVLM Overall Score RankVLM HiFi Removal Rank
LensErase1770.51465475911
ObjectClear1521.522903724434
OmniPaint1507.432793425822
OSOR1483.642564127543
RORem-mixed1462.552393829555
GeoRemover1410.662022234766
OmniEraser1343.971433339677


2. Sensitivity and potential artifact analysis of TraceFusion

📝 Question. How sensitive is TraceFusion to its heuristic parameters, and what potential issues or blending artifacts can occur?

💡 Response. Robustness in practice. TraceFusion is a simple, direct, and optional post-processing step. Because our RL-trained LensErase model already removes targets and associated effects cleanly while keeping background color and lighting consistent, the residual contrast between edited and unedited regions is sharp. Combined with median-bias subtraction (Eq. 27) and linear-RGB boundary blending (Eqs. 7 and 31), TraceFusion is insensitive to moderate parameter changes. Under our default settings (4/10 residual thresholds and 10% feather width), we observe no noticeable support-estimation errors, blending artifacts, or clear fusion failure cases across all 500 cases of LensErase-Bench or in regular usage. Moreover, since the raw output itself already looks natural, users can flexibly choose whether to apply TraceFusion.

Visual sensitivity analysis (Appendix K, Figures 22 to 25). In Appendix K of the revised manuscript, we vary the two core parameters in Appendix E.5 on fixed raw outputs to illustrate extreme regimes:

  1. Weak/strong residual thresholds (Figures 22 and 23, spanning 2/6 around our 4/10 default up to 64/128 on the 0 to 255 scale): Overly strict pairs (such as 18/36 or 64/128) narrow the estimated support and can restore faint shadows or reflections from the source image, whereas looser pairs (2/6) broaden the support and can retain minor background drift.
  2. Feather width (Figures 24 and 25, varying the 0.1 normalized distance in Eq. 31 across 0%, 5%, 10%, and 25%): Without feathering (0%), an abrupt switch to the source image can expose a subtle boundary seam when minor tone differences exist, whereas 5% to 10% feathering smoothly eliminates such transitions.

Across diverse scenes, our default choice (4/10 thresholds and 10% feather width) consistently avoids both issues.

3. Shared VLM and the difference between benchmark gains

📝 Question. Does using GPT-5.6 Sol as both the RL reward model and the LensErase-Bench evaluator create an evaluation loop that inflates the RL gain (70.80% to 81.60% HiFi Removal vs. 0.0826 to 0.0816 LPIPS on public benchmarks)?

💡 Response. First, the marginal LPIPS change on public benchmarks versus the large gain on LensErase-Bench comes from benchmark saturation and task difficulty: on some simpler public benchmarks, our SFT model already reaches a very low LPIPS, leaving little room or metric sensitivity for further gains. Second, the large HiFi Removal boost on LensErase-Bench reflects genuine improvements on hard cases, not VLM reward hacking. Please see our detailed analyses below:

1. Saturation on public benchmarks vs. headroom on LensErase-Bench. Some existing benchmarks mostly contain simpler edits where our SFT model is already near saturation. At an average LPIPS as low as 0.0826, errors are relatively minimal, leading to a small shift after RL. Even when evaluated under the exact same GPT-5.6 Sol rubric (before TraceFusion), SFT already reaches 87.73% HiFi Removal on OBER-Test and 86.76% on RemovalBench, yielding smaller RL gains (+3.68 and +5.89 pp). By contrast, LensErase-Bench targets much harder edits, including extended shadows or reflections, dense lookalike clusters and fine background textures. Starting from a lower SFT baseline (70.80%), RL has far greater headroom to resolve failed cases (+10.80 pp).

BenchmarkSFT OverallRL (Raw) OverallΔ OverallSFT HiFi (%)RL (Raw) HiFi (%)Δ HiFi (pp)
OBER-Test96.5097.09+0.5987.7391.41+3.68
RemovalBench96.1897.94+1.7686.7692.65+5.89
LensErase-Bench89.9194.41+4.5070.8081.60+10.80

2. SFT-to-RL gains across independent VLMs. We re-evaluated the exact same SFT and RL outputs (before TraceFusion) on LensErase-Bench using three unseen judges: GPT-6 Astra, GPT-6.1 Sol, and cross-family Gemini 3.8 Flash. All three independent judges confirm consistent gains of +3.64 to +4.38 points in Overall Score and +10.20 to +10.80 percentage points in HiFi Removal. These steady improvements confirm that RL genuinely resolves incomplete removals and local artifacts, without relying on judge-specific biases.

JudgeSFT OverallRL (Raw) OverallΔ OverallSFT HiFi Removal (%)RL (Raw) HiFi Removal (%)Δ HiFi Removal (pp)
GPT-5.6 Sol (paper)89.9194.41+4.5070.8081.60+10.80
GPT-6 Astra91.4595.09+3.6470.2081.00+10.80
GPT-6.1 Sol92.0395.87+3.8472.8083.00+10.20
Gemini 3.8 Flash92.3196.69+4.3879.4090.00+10.60

3. Method rankings across independent VLMs.

We re-evaluated LensErase and the six representative baselines across all 500 cases of LensErase-Bench with three independent VLM judges.

Across all four judges, LensErase consistently ranks 1st in both Overall Score and HiFi Removal ratio. Competing baselines also show only minor rank swaps among closely scored models, confirming the robustness of our evaluation rubric across VLM families.

OverallGPT-5.6 Sol ScoreGPT-5.6 Sol RankGPT-6 Astra ScoreGPT-6 Astra RankGPT-6.1 Sol ScoreGPT-6.1 Sol RankGemini 3.8 Flash ScoreGemini 3.8 Flash Rank
LensErase (w/ TraceFusion)95.22195.29195.61196.871
OmniPaint73.59273.91475.04477.465
OSOR71.45475.37276.66280.883
ObjectClear71.67374.06376.26381.182
RORem-mixed69.75572.63574.35579.964
GeoRemover62.70660.39761.92767.747
OmniEraser58.89763.97665.93669.726
HiFi Removal (%)GPT-5.6 Sol ScoreGPT-5.6 Sol RankGPT-6 Astra ScoreGPT-6 Astra RankGPT-6.1 Sol ScoreGPT-6.1 Sol RankGemini 3.8 Flash ScoreGemini 3.8 Flash Rank
LensErase (w/ TraceFusion)82.80181.00183.60190.201
OmniPaint42.80235.20238.20248.804
OSOR32.60333.20336.20354.402
ObjectClear28.40428.60434.00449.803
RORem-mixed27.40523.20526.00545.805
GeoRemover25.60620.80625.40637.406
OmniEraser22.60719.80723.60733.807


Summary. Taken together, these results address both aspects of the concern. First, the smaller gain on public benchmarks comes from near-saturated SFT performance on some simpler benchmarks, whereas the harder cases in LensErase-Bench leave much larger room for RL to resolve failures. Second, three unseen VLM judges reproduce the RL gain, and human judgments confirm our top ranking. This demonstrates that the RL gain reflects genuine improvements in removal quality instead of VLM reward hacking.



We have incorporated these results and clarifications into the revised manuscript.

I think my concerns are resolved and I keep my rating.

Thank you very much for your follow-up. We are pleased to hear that our explanations and additional experiments have addressed your concerns. We sincerely appreciate your constructive feedback.

Reviewer_sbHvOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

LensErase addresses image removal across objects, visual text, watermarks, poster elements, and cases with associated effects such as shadows or reflections. The model takes a source image and a selection mask. A useful distinction is that the mask identifies the content to remove, while the area needing modification may extend beyond it. Keeping the full source image also retains information that may be useful for reconstructing regions covered by semitransparent content. The paper introduces LensErase-100M, consisting of 101.07 million source–mask–target triplets from several task-specific data pipelines. For general objects, an object is inserted into a clean image, and the original image becomes the removal target. LensErase adapts the pretrained Lens model with two visual branches and Mask-Aware RoPE. Region-weighted flow-matching pretraining is followed by SFT and DiffusionNFT-based RL with candidate-independent removal contracts. At inference, TraceFusion estimates the edited region and restores source pixels outside it. The authors evaluate on six public benchmarks and the 500-case LensErase-Bench, compare with twelve baselines, and report stage-wise and component ablations. I view the main contribution as an effective, carefully assembled system for image removal, with the large-scale data construction and task-specific design working together.

Critical strengths
  1. The method is well matched to the practical challenges of image removal. The paper handles an important distinction between selecting an object and deciding which pixels should change. Preserving the full source image helps with translucent watermarks, while the model can remove related shadows or reflections outside the selection. I also appreciate the construction of clean supervision: the general-object pipeline uses an untouched real image as its target, and the watermark and poster pipelines produce aligned pairs through controlled synthesis or layer rendering. These choices address sources of noise that are especially relevant to removal tasks. Mask-Aware RoPE is a sensible addition to this design. At the pretraining stage, Table 4 reports 65.8% HiFi Removal for the proposed method, versus 58.8% for Channel Concat and 51.4% for Token Concat. The alternatives are not identical in every architectural detail, so I read this as support for the overall conditioning design rather than a pure positional-encoding effect. Even so, it is a useful task-specific result.
  2. The empirical results are strong, with useful evidence for the main training stages. The evaluation is substantial. On RORD-Val-343, LensErase improves PSNR from the strongest reported baseline's 26.21 to 28.23 dB; on TextEraseBench, it improves from 32.04 to 36.88 dB. The gains vary across datasets—for example, the improvement on the OmniPaint-Bench captured subset is 0.20 dB in PSNR—but the overall results across the six public benchmarks are convincing. The LensErase-Bench comparisons also show a clear advantage over the evaluated methods. The stage-wise results in Table 3 are helpful for understanding the pipeline. HiFi Removal rises from 70.8% after SFT to 81.6% after RL and reaches 82.8% with TraceFusion. TraceFusion has a larger effect on public-benchmark PSNR, increasing the average from 28.70 to 30.82 dB, which is consistent with better preservation outside the edited area. This separation of effects makes the results easier to interpret. Appendix G, Table 17 provides another useful check: jointly removing the Removal reward and its gate reduces HiFi Removal from 81.6% to 38.8%, even though the mean Preservation score reaches 98.7/100. This illustrates why preservation alone is not enough for the task. Overall, the evidence supports a meaningful contribution, and I lean toward weak acceptance despite the relatively modest concerns below.
Critical weaknesses
  1. The semantic evaluation would benefit from a small independent check. GPT-5.6 Sol is involved in SFT sample filtering, RL reward evaluation, and the LensErase-Bench evaluation. The reward and benchmark protocols are different, which is useful, but shared judging preferences may still affect the reported semantic scores. The 82.8% HiFi Removal result is clearly defined under the paper's VLM rubric; a small human spot check or evaluation with an independent judge would help show whether the method ranking and the SFT-to-RL gain hold under another assessment. The benchmark protocol also uses one output per case with a fixed, case-specific seed. This is appropriate for a controlled comparison, although it does not tell us how much performance varies for the same input across different seeds. I see these as limitations on the interpretation of the semantic results, not as evidence against the strong public-benchmark improvements. A brief discussion of judge dependence, together with any existing spot-check results or representative failure cases, would be sufficient for my present assessment; I would not require a large new evaluation.
  2. A few details of the ablations and dataset separation could be clearer. The ablations support the combined method, but they do not isolate every component. In Table 17, the Removal reward and removal-aware gate are disabled together, so the result does not measure the gate's separate contribution. Likewise, Token Concat changes branch routing as well as mask representation. These are mainly attribution limits; the comparisons, especially Channel Concat, are still useful evidence for the proposed design. The paper also provides helpful initialization and training-schedule comparisons in Appendix D. A data-scaling study could offer more detail about the role of the very large training set, but I do not think it is essential to establish the value of the end-to-end system presented here. For dataset separation, the paper states that SFT and RL use disjoint source images and describes near-duplicate suppression within RL. I would appreciate a short clarification of how the six public evaluation sets and LensErase-Bench were checked against the training data, including the internal evaluation used to guide failure-focused RL data collection. This would make the evaluation protocol easier to assess; it is a request for documentation rather than an allegation of leakage. These points seem addressable through clarification and a more precise statement of what the available ablations establish. They do not substantially change my positive overall assessment.
Questions for authors
  1. Independent evaluation. Have the authors carried out any human spot checks or evaluations with a different VLM judge on a small subset of LensErase-Bench? In particular, does the SFT-to-RL improvement look similar? If such results are unavailable, a short discussion of potential judge dependence and a few representative failure cases would help.
  2. Evaluation-set separation. Could the authors briefly explain how the six public benchmarks and LensErase-Bench were kept separate from pretraining, SFT, and RL data, including the internal evaluation that informed failure-focused RL collection? A concise description of the image-level and near-duplicate checks would be sufficient.
  3. Interpreting the ablations. Could the authors clarify that Table 17 supports the Removal reward and gate jointly rather than isolating the gate? If a gate-only result is already available, it would be interesting to see, but I do not consider another large training run necessary.
5 comments

We sincerely thank Reviewer sbHv for the thoughtful assessment and constructive suggestions. Below, we address each point with new multi-judge and multi-seed evaluations, a 2,000-judgment human pairwise study, deduplication details, an isolated gate ablation, and a data-scale study.

Please feel free to let us know if any further discussions or experiments would be helpful.



1. Method rankings across VLM judges

📝 Question. Do alternative VLM judges reproduce the method ranking?

💡 Response. Yes. Using the exact same model outputs and scoring rubric as in Table 2, we re-evaluated LensErase and the six representative baselines across all 500 cases of LensErase-Bench with three independent VLM judges, including two newer models (GPT-6 Astra, GPT-6.1 Sol) and a cross-family judge (Gemini 3.8 Flash).

Across all four judges, LensErase consistently ranks 1st in both Overall Score and HiFi Removal ratio. Competing baselines also show only minor rank swaps among closely scored models, confirming the robustness of our evaluation rubric across VLM families.

Method · Overall ScoreGPT-5.6 Sol ScoreGPT-5.6 Sol RankGPT-6 Astra ScoreGPT-6 Astra RankGPT-6.1 Sol ScoreGPT-6.1 Sol RankGemini 3.8 Flash ScoreGemini 3.8 Flash Rank
LensErase (w/ TraceFusion)95.22195.29195.61196.871
OmniPaint73.59273.91475.04477.465
OSOR71.45475.37276.66280.883
ObjectClear71.67374.06376.26381.182
RORem-mixed69.75572.63574.35579.964
GeoRemover62.70660.39761.92767.747
OmniEraser58.89763.97665.93669.726
Method · HiFi Removal (%)GPT-5.6 Sol ScoreGPT-5.6 Sol RankGPT-6 Astra ScoreGPT-6 Astra RankGPT-6.1 Sol ScoreGPT-6.1 Sol RankGemini 3.8 Flash ScoreGemini 3.8 Flash Rank
LensErase (w/ TraceFusion)82.80181.00183.60190.201
OmniPaint42.80235.20238.20248.804
OSOR32.60333.20336.20354.402
ObjectClear28.40428.60434.00449.803
RORem-mixed27.40523.20526.00545.805
GeoRemover25.60620.80625.40637.406
OmniEraser22.60719.80723.60733.807

2. Human pairwise evaluation

📝 Question. Do human preferences support the method ranking?

💡 Response. Yes. We conducted a blind human preference study with 20 participants, each completing 100 pairwise comparisons across LensErase and the six baselines in Table 2 (2,000 judgments in total). In each trial, participants viewed the source image with the selected region indicated alongside two anonymized outputs in randomized A/B order, and selected the output that more completely removed the target and its associated effects, reconstructed the region more plausibly, and better preserved non-target content (with ties allowed).

We fit a Bradley–Terry model [1], counting ties as half-wins and centering ratings at 1500 Elo. LensErase ranks first with 1770.5 Elo, outperforming the strongest baseline (ObjectClear) by 249.0 points. It also wins 81.4% of its matchups outright. Moreover, the human ranking closely aligns with the VLM-based HiFi Removal rankings. The same three baselines occupy ranks 2–4, while ranks 5–7 match exactly, providing independent human validation of our VLM evaluation protocol.

[1] Bradley, Ralph Allan, and Milton E. Terry. "Rank analysis of incomplete block designs: I. the method of paired comparisons." Biometrika 39.3/4 (1952): 324-345.

MethodElo ↑Human RankWinsTiesLossesVLM Overall Score RankVLM HiFi Removal Rank
LensErase1770.51465475911
ObjectClear1521.522903724434
OmniPaint1507.432793425822
OSOR1483.642564127543
RORem-mixed1462.552393829555
GeoRemover1410.662022234766
OmniEraser1343.971433339677


3. SFT-to-RL gain across VLM judges

📝 Question. Does the improvement from SFT to RL hold under independent VLM judges?

💡 Response. Yes. We evaluated the SFT and RL checkpoints (both before TraceFusion) across all 500 LensErase-Bench cases under the three independent judges. Although the judges differ in absolute score calibration, all three independent judges confirm substantial and remarkably consistent SFT-to-RL gains: +3.64 to +4.38 points in Overall Score and +10.20 to +10.80 percentage points in HiFi Removal. This calibration-invariant improvement confirms that RL genuinely resolves incomplete removals and local artifacts, without relying on judge-specific biases.

JudgeSFT OverallRL (Raw) OverallΔ OverallSFT HiFi Removal (%)RL (Raw) HiFi Removal (%)Δ HiFi Removal (pp)
GPT-5.6 Sol (paper)89.9194.41+4.5070.8081.60+10.80
GPT-6 Astra91.4595.09+3.6470.2081.00+10.80
GPT-6.1 Sol92.0395.87+3.8472.8083.00+10.20
Gemini 3.8 Flash92.3196.69+4.3879.4090.00+10.60

4. Inference seeds

📝 Question. How much do the results vary across inference seeds?

💡 Response. To measure cross-seed variance on identical inputs, we re-evaluated the full LensErase model (with TraceFusion) across all 500 cases of LensErase-Bench under three independent fixed seeds (20921, 27189, and 57327) using the GPT-5.6 Sol judge. Across the three runs, LensErase achieves 95.13 ± 0.10 in Overall Score and 82.00% ± 0.40% in HiFi Removal (closely matching the 95.22 Overall and 82.80% HiFi Removal reported in Table 2). This confirms that our RL-trained model varied little across the three tested seeds.

SeedRemovalInpaintingPreservationCleanlinessOverallHiFi Removal (%)
2092195.2592.0096.0597.1595.1482.40
2718995.0591.6096.4597.0095.0381.60
5732795.1591.9096.1097.8095.2282.00


5. Evaluation-set separation

📝 Question. What exact-image and near-duplicate checks were used for the six public benchmarks and LensErase-Bench against pretraining, SFT, and RL, including the internal evaluation that guided RL data collection?

💡 Response. Source-level split and near-duplicate filtering. Before synthesis or annotation, we partitioned the original images into four mutually exclusive source pools for pretraining, SFT, RL, and LensErase-Bench. Each original image was assigned to only one pool, including images drawn from shared public collections such as layered poster designs and segmentation datasets. We kept the source images of the six public benchmarks separate from all training pools. We also computed a 64-bit perceptual hash (pHash) for each SFT and RL candidate source image and each evaluation source image. We excluded any SFT or RL candidate whose hash differed by at most four bits from that of any source image in LensErase-Bench or the six public benchmarks.

Internal diagnostic evaluation and RL data collection. None of the seven benchmarks were used during internal diagnostics. Instead, we evaluated LensErase-SFT in a separate internal trial. Internal users submitted everyday image-editing requests on independent images and marked unsatisfactory results. Reviewing these failed cases revealed several recurring failure modes. To address them, we combined manual curation with GPT-5.6 Sol to screen candidate images and construct targeted RL tasks (with SAM2 refining object masks where needed). Representative examples include:

  1. Residual shadows and reflections: Using GPT-5.6 Sol to mine images with prominent shadows or reflections, and masking only the casting objects.

  2. Over-erasure in dense clusters: Masking a subset of closely spaced, similar objects so that neighboring instances had to be preserved.

  3. Blurry reconstruction of fine textures: Selecting foreground targets that occluded finely textured backgrounds.

All newly collected RL images also passed the same 64-bit pHash check (Hamming distance <= 4) against SFT and all seven benchmarks.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair, Senior Area Chairs, and Reviewers,

We sincerely thank you for your careful assessment and constructive feedback. Your comments have helped us strengthen the experimental evaluation, clarify methodological details, and improve the presentation of our work.

Below, we summarize the main concerns and the corresponding results and clarifications provided in our responses and revision.

Main ConcernReviewersAdditional Results and Clarifications
Human preferencessbHv, 15uf, 2eszLensErase ranked first in a blind human preference study involving 2,000 judgments, winning 81.4% of its matchups outright.
Method ranking across VLM judgessbHv, 15uf, 2eszLensErase ranked first among seven methods in both Overall Score and HiFi Removal under all four VLM judges.
SFT-to-RL gains under independent judgessbHv, 15uf, 2eszAll three additional VLM judges showed consistent SFT-to-RL improvements in Overall Score and HiFi Removal across 500 LensErase-Bench cases.
Contribution of pretraining scalesbHv, 2eszScaling pretraining from 500K to 101.07M triplets improved HiFi Removal from 71.2% to 82.8%, while the 500K variant already substantially outperformed prior SOTA methods.
Effect of the removal-aware gatesbHv, 2eszWith the Removal reward held fixed, the gate mechanism improved HiFi Removal by 4.8 percentage points on LensErase-Bench and mean PSNR by 0.51 dB across six public benchmarks.
TraceFusion sensitivity15ufWe examined different residual thresholds and feathering widths, characterizing their effects on faint shadow reappearance and boundary artifacts.
Inference seed stability and data separationsbHvHiFi Removal ranged from 81.6% to 82.8% across three seeds. We further clarified the source-image splits and near-duplicate screening of SFT and RL candidates against all seven benchmarks.

Thank you again for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.