LensErase: A Foundation Model for Erase Anything

Yunuo Chen; Xiuyu Wu; Qinhong Yang; Jinjing Zhao; Yanchen Dong; Sirui Zhang; Dong Chen; Fangyun Wei

cs.CVOct 13, 2026

LensErase: A Foundation Model for Erase Anything

Authors: Yunuo Chen, Xiuyu Wu, Qinhong Yang, Jinjing Zhao, Yanchen Dong, Sirui Zhang, Dong Chen, Fangyun Wei,

Organizations: Shanghai Jiao Tong University · Nankai University · USTC · University of Sydney ·

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.

Figures & tables

Figure 1: LensErase across diverse removal scenarios. Each pair shows a source image with highlighted removal targets and the corresponding LensErase result. Examples cover general objects, visual text, watermarks, multiple objects, coupled effects, and high-fidelity inpainting.

Figure 2: Left : Distribution of mask area ratios. Right : Distribution of mask counts.

Figure 3: Architecture of LensErase. LensErase follows a two-branch MMDiT architecture, where the image branch processes the noisy target latent and the condition branch processes the source latent. The proposed Mask-Aware RoPE incorporates the pooled binary mask into the condition branch by encoding masked and unmasked source regions along an additional modality dimension.

Method	PSNR $\uparrow$	FID $\downarrow$	CMMD $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	DISTS $\downarrow$	CLIP-I $\uparrow$	DINO $\uparrow$
OBER-Test [ 75 ]
ObjectClear [ 75 ]	33.05	19.16	0.0132	0.0338	0.9353	0.0217	0.9894	0.9839
OSOR [ 78 ]	30.14	27.82	0.0674	0.0524	0.8748	0.0377	0.9827	0.9693
OmniPaint [ 71 ]	28.79	23.68	0.0829	0.0549	0.8575	0.0427	0.9836	0.9803
LensErase	34.96	16.97	0.0073	0.0256	0.9543	0.0176	0.9930	0.9839
RORD-Val-343 [ 50 ]

Table 1: Comparison with state-of-the-art object removal methods on six public benchmarks. For each benchmark, we report the three baselines with the lowest mean rank across the eight metrics together with LensErase . Complete results for all 12 baselines are provided in Tables 10 - 15 of the Appendix. Cell backgrounds mark the best , second-best , and third-best scores among these four methods.

Method	Removal $\uparrow$	Inpainting $\uparrow$	Preservation $\uparrow$	Cleanliness $\uparrow$	Overall $\uparrow$	HiFi Removal (%) $\uparrow$
OSOR [ 78 ]	73.85	60.45	75.55	73.55	71.45	32.60
OmniPaint [ 71 ]	69.20	66.15	80.20	83.20	73.59	42.80
ObjectClear [ 75 ]	73.95	61.45	73.80	75.20	71.67	28.40
GeoRemover [ 80 ]	54.30	49.85	77.05	78.00	62.70	25.60
OmniEraser [ 62 ]	55.65	47.95	77.55	57.65	58.89	22.60
RORem-mixed [ 31 ]	73.95	59.70	74.05	67.10	69.75	27.40

Table 2: Comparison on LensErase-Bench. We report the six representative baselines together with LensErase. Complete results for all 12 baselines are provided in Table 16 of the Appendix.

	LensErase-Bench						Public Benchmarks
Model	Removal $\uparrow$	Inpainting $\uparrow$	Preservation $\uparrow$	Cleanliness $\uparrow$	Overall $\uparrow$	HiFi Removal $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
Pre-Training	82.35	84.65	91.05	92.15	86.51	65.80	28.03	0.0855
+SFT	88.65	87.15	92.05	93.05	89.91	70.80	28.14	0.0826
+RL	94.95	92.65	93.35	96.15	94.41	81.60	28.70	0.0816
+TraceFusion	95.15	92.40	96.35	97.05	95.22	82.80	30.82	0.0684

Table 3: Stage-wise ablation of LensErase. We report performance after pre-training, SFT, RL, and TraceFusion. Results on LensErase-Bench are shown with all evaluation metrics, while results on public benchmarks report the average PSNR and LPIPS across the six public benchmarks.

	LensErase-Bench						Public Benchmarks
Model	Removal $\uparrow$	Inpainting $\uparrow$	Preservation $\uparrow$	Cleanliness $\uparrow$	Overall $\uparrow$	HiFi Removal $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
Mask-Aware RoPE	82.35	84.65	91.05	92.15	86.51	65.80	28.03	0.0855
Channel Concat	81.50	84.55	82.90	91.15	84.32	58.80	27.84	0.0869
Token Concat	77.50	76.85	84.90	87.15	80.78	51.40	27.41	0.0965

Table 4: Comparison of mask injection strategies. Mask-Aware RoPE is compared with channel- and token-concatenation baselines, all evaluated at the pre-training stage. Both baselines obtain mask features via pixel unshuffle of

16\times 16

patches. Channel Concat concatenates source and mask features along channels and fuses them with an MLP before condition projection. Token Concat places source tokens alongside noisy tokens in the image branch and projects mask features to the condition branch. Both use positional indices independent of mask values.

Appendix figures & tables31 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Data composition across training stages. Dataset sizes and category counts for pretraining, SFT, and RL. Bands connect shared categories across stages.

Domain	Mask-area bins (%)	Cases per bin
General	$<1,\ [1,5),\ [5,15),\ [15,30),\ \geq 30$	40 / 40 / 40 / 40 / 40
Poster	$<5,\ [5,15),\ \geq 15$	35 / 33 / 32
Watermark	$<1,\ [1,5),\ [5,15),\ \geq 15$	25 / 25 / 25 / 25
Visual Text	$<0.5,\ [0.5,1),\ [1,2),\ \geq 2$	25 / 25 / 25 / 25

Appendix

Table 5: Domain-specific mask-area distribution in LensErase-Bench.

	Filling difficulty				Selected targets
Domain	0–1	2	3	4	1	2–3	$\geq 4$
General	9	50	87	54	100	87	13
Poster	34	25	25	16	93	3	4
Watermark	0	9	64	27	40	16	44
Visual Text	25	45	30	0	63	25	12

Appendix

Table 6: Distribution of filling difficulty and number of selected targets in LensErase-Bench.

Figure 5: Initialization and two-stage pretraining. Both visual branches are initialized from the pretrained Lens image branch. Stage 1 adapts the condition branch while freezing the image-specific backbone. Stage 2 jointly adapts both branches for removal and reconstruction.

Conditioning and initialization
	LensErase-Bench						Public Benchmarks
Model	Removal $\uparrow$	Inpainting $\uparrow$	Preservation $\uparrow$	Cleanliness $\uparrow$	Overall $\uparrow$	HiFi Removal $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
Mask-Aware RoPE (Stage 2)	82.35	84.65	91.05	92.15	86.51	65.80	28.03	0.0855
Channel Concat	81.50	84.55	82.90	91.15	84.32	58.80	27.84	0.0869
Token Concat	77.50	76.85	84.90	87.15	80.78	51.40	27.41	0.0965
Text initialization	89.50	1.30	0.05	1.75	36.42	0.00	10.71	0.7420

Appendix

Table 7: Pretraining design ablations. The shared reference uses Mask-Aware RoPE, image-branch initialization, and two-stage freeze–unfreeze training. All rows use raw outputs. LensErase-Bench metrics follow Section 2.3 ; PSNR and LPIPS are equal-weight means over the six public benchmarks.

Figure 6: Contract-guided reward construction for General Data. Source analysis establishes an erase contract shared across rollouts. Each candidate is evaluated against this fixed contract using semantic judgments and pixel fidelity. The five reward components are combined and passed through the removal-aware gate in Equation 6 . The resulting rewards define group advantages for DiffusionNFT.

Component	General Data	Poster / Watermark / Visual Text	General Data criteria
Removal	0.30	0.20	Target body absence (0.35), parts/identity absence (0.30), associated-effect removal (0.35)
Reconstruction	0.15	0.20	Semantic plausibility, structural continuity, and appearance/material consistency
Preservation	0.15	0.15	Independent-object preservation (0.60), background/scene-structure preservation (0.40)
Outside Fidelity	0.30	0.25	Outside-region PSNR score (1.00)
Cleanliness	0.10	0.20	Absence of residual fragments, unwanted additions (objects, duplicates, text, logos, or watermarks), and technical artifacts

Appendix

Table 8: Reward weights before applicability renormalization and removal-aware gating. The last column summarizes the General Data criteria. Poster, Watermark, and Visual Text share component weights, with reference-specific subcriteria described below.

Symbol	Definition
Images, masks, and positional coordinates
$I_{\text{input}}$	Source image containing the selected targets.
$I_{\text{output}}$	Paired erase target used for supervised training.
$\hat{I}_{\text{output}}$ , $\hat{I}_{\text{output}}^{\mathrm{TF}}$	Generated image before and after TraceFusion, respectively.
$M$ , $m_{i}$	Binary selection mask and its max-pooled value for source token $i$ , with 1 denoting a selected region.
$h_{i}$ , $w_{i}$	Height and width coordinates of token $i$ .

Appendix

Table 9: Principal notation for LensErase.

Method	PSNR $\uparrow$	FID $\downarrow$	CMMD $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	DISTS $\downarrow$	CLIP-I $\uparrow$	DINO $\uparrow$
SmartEraser [ 25 ]	26.67	94.58	0.4178	0.0762	0.9446	0.0792	0.9393	0.8423
RORem-mixed [ 31 ]	25.93	61.32	0.1832	0.1122	0.8549	0.0836	0.9593	0.9186
EntityErasure [ 79 ]	22.75	82.02	0.2933	0.1330	0.7263	0.1037	0.9376	0.8866
Attentive Eraser [ 55 ]	26.22	59.96	0.2048	0.1123	0.8497	0.0889	0.9605	0.9228
DesignEdit [ 24 ]	26.39	71.25	0.2052	0.1302	0.8585	0.0955	0.9485	0.8997
OmniEraser [ 62 ]	25.52	85.73	0.2131	0.0996	0.9141	0.0896	0.9457	0.8698

Appendix

Table 10: Complete comparison on OBER-Test [ 75 ] .

Method	PSNR $\uparrow$	FID $\downarrow$	CMMD $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	DISTS $\downarrow$	CLIP-I $\uparrow$	DINO $\uparrow$
SmartEraser [ 25 ]	24.89	35.41	0.1104	0.1658	0.8472	0.0950	0.9644	0.9349
RORem-mixed [ 31 ]	23.16	35.91	0.0983	0.1906	0.6893	0.0990	0.9482	0.9316
EntityErasure [ 79 ]	21.25	50.71	0.2829	0.3340	0.6142	0.1461	0.9400	0.9082
Attentive Eraser [ 55 ]	22.57	39.19	0.0948	0.1743	0.6896	0.0879	0.9501	0.9189
DesignEdit [ 24 ]	23.28	49.55	0.1603	0.1938	0.7268	0.1046	0.9304	0.8837
OmniEraser [ 62 ]	24.46	35.78	0.0426	0.1189	0.8380	0.0690	0.9580	0.9262

Appendix

Table 11: Complete comparison on RORD-Val-343 [ 50 ] .

Method	PSNR $\uparrow$	FID $\downarrow$	CMMD $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	DISTS $\downarrow$	CLIP-I $\uparrow$	DINO $\uparrow$
SmartEraser [ 25 ]	23.49	151.88	0.4359	0.3308	0.7695	0.1890	0.9339	0.8239
RORem-mixed [ 31 ]	23.51	90.95	0.2044	0.2262	0.7149	0.1175	0.9575	0.9098
EntityErasure [ 79 ]	20.36	161.80	0.3525	0.4287	0.6664	0.2288	0.9116	0.7918
Attentive Eraser [ 55 ]	23.78	80.49	0.1666	0.1941	0.7095	0.0969	0.9646	0.9263
DesignEdit [ 24 ]	23.96	108.46	0.2758	0.2343	0.7200	0.1357	0.9457	0.8676
OmniEraser [ 62 ]	22.33	147.90	0.3822	0.2150	0.7539	0.1285	0.9213	0.7990

Appendix

Table 12: Complete comparison on RemovalBench [ 62 ] .

Method	PSNR $\uparrow$	FID $\downarrow$	CMMD $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	DISTS $\downarrow$	CLIP-I $\uparrow$	DINO $\uparrow$
SmartEraser [ 25 ]	22.05	94.79	0.2854	0.2787	0.8121	0.1460	0.9189	0.8610
RORem-mixed [ 31 ]	23.14	54.66	0.1398	0.1994	0.7730	0.1072	0.9433	0.9226
EntityErasure [ 79 ]	20.44	96.41	0.3472	0.3033	0.7262	0.1583	0.9068	0.8579
Attentive Eraser [ 55 ]	23.31	50.60	0.1324	0.1642	0.7865	0.0719	0.9452	0.9322
DesignEdit [ 24 ]	23.30	68.54	0.2245	0.1838	0.7859	0.1033	0.9261	0.8838
OmniEraser [ 62 ]	21.73	93.71	0.2645	0.1767	0.8123	0.0904	0.9059	0.8212

Appendix

Table 13: Complete comparison on OmniPaint-Bench [ 71 ] (captured subset), excluding RORD-derived cases.

Method	PSNR $\uparrow$	FID $\downarrow$	CMMD $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	DISTS $\downarrow$	CLIP-I $\uparrow$	DINO $\uparrow$
SmartEraser [ 25 ]	23.90	55.01	0.0451	0.2116	0.8252	0.1174	0.9837	0.9577
RORem-mixed [ 31 ]	24.42	33.68	0.0261	0.1305	0.7926	0.0654	0.9875	0.9756
EntityErasure [ 79 ]	20.26	83.39	0.0906	0.3574	0.6143	0.1776	0.9711	0.9282
Attentive Eraser [ 55 ]	22.60	37.77	0.0315	0.1329	0.7321	0.0629	0.9852	0.9699
DesignEdit [ 24 ]	24.10	49.33	0.0666	0.1582	0.7861	0.0846	0.9800	0.9519
OmniEraser [ 62 ]	24.05	49.39	0.0383	0.1155	0.8569	0.0617	0.9772	0.9273

Appendix

Table 14: Complete comparison on AnimeEraseBench [ 78 ] .

Method	PSNR $\uparrow$	FID $\downarrow$	CMMD $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	DISTS $\downarrow$	CLIP-I $\uparrow$	DINO $\uparrow$
SmartEraser [ 25 ]	28.85	47.96	0.0608	0.1686	0.8885	0.1049	0.9430	0.9471
RORem-mixed [ 31 ]	29.50	24.83	0.0178	0.1034	0.8292	0.0584	0.9815	0.9744
EntityErasure [ 79 ]	24.98	53.79	0.1175	0.2653	0.7307	0.1502	0.9427	0.9414
Attentive Eraser [ 55 ]	26.31	34.70	0.0228	0.1134	0.7956	0.0563	0.9738	0.9613
DesignEdit [ 24 ]	27.61	49.23	0.1002	0.1391	0.8289	0.0883	0.9578	0.9247
OmniEraser [ 62 ]	30.70	21.97	0.0195	0.0763	0.9005	0.0435	0.9812	0.9758

Appendix

Table 15: Complete comparison on TextEraseBench [ 78 ] .

Method	Removal $\uparrow$	Inpainting $\uparrow$	Preservation $\uparrow$	Cleanliness $\uparrow$	Overall $\uparrow$	HiFi Removal (%) $\uparrow$
SmartEraser [ 25 ]	23.40	24.15	66.25	57.75	38.99	3.60
RORem-mixed [ 31 ]	73.95	59.70	74.05	67.10	69.75	27.40
EntityErasure [ 79 ]	59.65	48.50	48.45	60.20	55.29	4.80
Attentive Eraser [ 55 ]	78.40	52.25	61.50	55.05	65.12	20.60
DesignEdit [ 24 ]	76.95	60.25	63.85	62.95	68.19	25.00
OmniEraser [ 62 ]	55.65	47.95	77.55	57.65	58.89	22.60

Appendix

Table 16: VLM comparison on LensErase-Bench.

	LensErase-Bench						Public Benchmarks
Model	Removal $\uparrow$	Inpainting $\uparrow$	Preservation $\uparrow$	Cleanliness $\uparrow$	Overall $\uparrow$	HiFi Removal $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$
RL-Raw (full reward)	94.95	92.65	93.35	96.15	94.41	81.60	28.70	0.0816
w/o Removal + gate	43.35	46.10	98.70	98.10	65.92	38.80	21.85	0.1446
w/o Reconstruction	90.85	91.40	96.10	95.70	92.98	78.00	28.32	0.0791
w/o Preservation	94.15	91.50	89.20	96.00	93.00	74.60	28.01	0.0817
w/o Outside Fidelity	92.90	91.70	94.40	96.40	93.66	78.00	26.69	0.0888

Appendix

Table 17: Reward-component ablations. PSNR and LPIPS are equal-weight means over the six public benchmarks.

Figure 7: Qualitative comparisons with GPT-Image 2.5 and Flux2-Erase.

Figure 8: Qualitative comparisons with GPT-Image 2.5 and Flux2-Erase.

Figure 9: Qualitative comparisons with GPT-Image 2.5 and Flux2-Erase.

Figure 10: Qualitative comparisons with GPT-Image 2.5 and Flux2-Erase.

Figure 11: Qualitative comparisons on the General subset.

Figure 12: Qualitative comparisons on General images and posters.

Figure 13: Qualitative comparisons on posters.

Figure 14: Qualitative comparisons on posters.

Figure 15: Qualitative comparisons on watermark removal.

Figure 16: Qualitative comparisons on watermark and visual text removal.

Figure 17: Qualitative comparisons on visual text removal.

Figure 18: LensErase results with user-provided masks.

Figure 19: LensErase results with user-provided masks.

Figure 20: LensErase results with user-provided masks.

Figure 21: LensErase results with user-provided masks.

Image Editing

Image editing changes selected visual content while preserving the parts of an image that should remain unchanged.

See all 302 papers

EraseLoRA: MLLM-Driven Foreground Exclusion and Background Subtype Aggregation for Dataset-Free Object Removal
Dec 25, 2025Sanghyun Jo, Donghwan Lee, Eunji Jung +2Object Removal Diffusion-Based Image Editing
OGQ, Seoul, Korea · Seoul National University, Seoul, Korea
You Only Erase Once: Erasing Anything without Bringing Unexpected Content
Mar 29, 2026Yixing Zhu, Qing Zhang, Wenju Xu +1Object Removal Diffusion-Based Image Editing
School of Computer Science and Engineering, Sun Yat-sen University, China · Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China · Amazon
GenEraser: Generalizable Video Object Removal via Balanced Text-Mask Guidance and Decoupled Locator-Preserver
May 28, 2026Yuqing Chen, Lin Liu, Haisu Wu +4Diffusion Model Guidance Video Diffusion Models
Tsinghua University, China · Huawei, China · Southeast University, China +2

Dec 25, 2025J/K move · Enter open · S save

EraseLoRA: MLLM-Driven Foreground Exclusion and Background Subtype Aggregation for Dataset-Free Object Removal

Sanghyun Jo, Donghwan Lee, Eunji Jung, Seong Je Oh+1

OGQ, Seoul, Korea · Seoul National University, Seoul, Korea

Object Removal Diffusion-Based Image Editing

Open paper AskPDF arXiv

Object removal must prevent the masked target from reappearing and reconstruct the occluded background with structural and contextual fidelity, rather than merely filling a hole plausibly. Recent dataset-free approaches manipulate the diffusion model's internal self-attention to prevent it from referencing the masked region, yet they fail in two critical ways: (i) they treat the masked region as the sole foreground, misinterpreting non-target objects as background and regenerating them, and (ii) they apply uniform attention constraints without distinguishing diverse background subtypes, leading to textural blurring and structural misalignment. Both failures stem from the absence of explicit background-aware reasoning. We propose EraseLoRA, a dataset-free framework that replaces attention surgery with background-aware reasoning and test-time adaptation. The first stage, Background-aware Foreground Exclusion (BFE), leverages a multimodal large-language model to separate target foreground, non-target foregrounds, and clean background from a single image-mask pair. The second stage, Background-aware Reconstruction with Subtype Aggregation (BRSA), performs test-time optimization that treats inferred background subtypes as complementary pieces, enforcing their consistent integration through reconstruction and alignment objectives without explicit attention intervention. As a model-agnostic plug-in applicable to diverse diffusion backbones, EraseLoRA reconstructs backgrounds at least 23% more faithful to the original scene than previous dataset-free methods while nearly halving unwanted foreground re-generation, and surpasses all dataset-driven approaches in both aspects despite requiring no training data. Code is available at https://shjo-april.github.io/EraseLoRA.

Abstract

Image Editing

Explore similar work

EraseLoRA: MLLM-Driven Foreground Exclusion and Background Subtype Aggregation for Dataset-Free Object Removal

You Only Erase Once: Erasing Anything without Bringing Unexpected Content

GenEraser: Generalizable Video Object Removal via Balanced Text-Mask Guidance and Decoupled Locator-Preserver