Object removal and semantic inpainting of missing regions remain inherently ill-posed, as multiple plausible completions may exist for a given scene. Recent diffusion-based methods have demonstrated strong potential for high-quality object removal; however, they still face two key limitations. First, although they can synthesise sharp and high-resolution images, they often struggle to complete removed regions in a semantically consistent and contextually meaningful manner, where the generated content should agree with the surrounding scene layout and object relationships rather than merely appear visually realistic. Second, their iterative denoising process makes them computationally expensive, limiting their practicality for downstream applications. We propose SERO, a semantic prediction framework for single-step object removal. Our model predicts a hidden-region semantic prior in embedding space, which guides a latent bridge-matching generator toward plausible scene completion by providing both where to edit and what structure is likely to exist behind the removed object. This enables visually realistic and contextually coherent completions in one forward pass. Experiments on the RORD and OBER datasets show that SERO achieves competitive or state-of-the-art performance while requiring only a single generation step.
Figures & tables
Figure 1. SERO combines semantic prediction with efficient bridge matching to remove objects and reconstruct plausible scene content in a single step, across both real-world object removal and complex scene-completion scenarios. An overview and visual examples of the SERO single-step object-removal method.
Figure 2. Overview of the proposed pipeline. (a) Training: the frozen context encoder and predictor produce hidden-region token embeddings from visible background patches, giving us a spatial prior; a trainable decoder adapter maps this to a dense latent prior that is concatenated with the bridge latent and mask before being fed to the Denoising U-Net. (b) Inference: given a text prompt and input image, a segmentation model generates the removal mask, after which the semantic-prior-conditioned SERO model predicts the clean output in a single forward pass. The SERO training and inference pipelines. During training, a frozen context encoder and predictor estimate hidden-region embeddings, an adaptor maps them to a spatial prior, and a U-Net learns object removal. During inference, a prompt-derived mask and the semantic prior guide a single-step clean-image prediction.
RORD-Val
OBER-Test
Method
FID ↓
localFID ↓
fMSE ↓
PSNR FG↑
LPIPS ↓
CLIP ↓
FID ↓
localFID ↓
fMSE ↓
PSNR FG↑
LPIPS ↓
CLIP ↓
PowerPaint (ECCV’24)
53.81
65.34
2508.69
15.84
0.2346
0.0715
67.37
90.98
1916.73
18.62
0.1243
0.0409
SDXL (ICLR’24)
67.18
79.72
2947.48
14.74
0.2308
0.1101
75.44
109.50
2573.60
16.15
0.1296
0.0579
OmniEraser (arXiv’25)
46.80
51.65
1247.13
18.62
0.1962
0.0536
54.70
54.86
883.91
21.00
0.1783
0.0142
GeoRemover (NeurIPS’25)
27.77
42.28
792.00
21.18
0.1009
0.0320
50.08
50.80
769.53
22.12
0.1976
0.0670
LBM (ICCV’25)
43.01
60.37
1301.02
18.45
0.1768
0.0470
29.13
54.20
477.60
23.75
0.0824
0.0171
Table 1. Quantitative results on RORD-Val and OBER-Test. We report FID, localFID, fMSE, PSNR FG , LPIPS, and CLIP distance across both benchmark datasets. Bold denotes best and underline denotes second best per column.
Figure 3. Qualitative comparison on RORD-Val. Each row corresponds to one method and each column to one scene. Compared with PowerPaint, OmniEraser, GeoRemover, and ObjectClear, SERO removes the target object while producing the most semantically coherent and artifact-free background completion, closely matching the ground truth. Four RORD-Val scenes arranged as columns, with one row per method: the masked input, four prior methods, SERO, and the ground-truth background.
Figure 4. Qualitative comparison on OBER-Test. Each column corresponds to one method and each row to one scene. In addition to the object itself, SERO removes the associated visual effects, such as the reflection in the first row, while preserving the surrounding scene; prior methods either leave residual traces of the effect or alter content outside the mask. Three OBER-Test scenes arranged as rows, with one column per method: the masked input, four prior methods, SERO, and the ground-truth background.
Figure 5. Quality–efficiency comparison on the RORD-Val dataset. FID is plotted against the number of denoising steps; lower is better for both axes. SERO achieves the best trade-off, attaining the lowest FID with only a single denoising step. Scatter plot of FID versus denoising steps. SERO appears at one step with the lowest FID, while competing methods use between 20 and 100 steps.
Method
Steps
Median (s)
Slowdown
Trainable (B)
LBM
4
0.530
1.7×
2.43
ObjectClear
20
1.442
4.5×
3.79
SDXL-Inpainting
50
2.342
7.3×
3.47
PowerPaint v2
45
6.549
20.4×
2.08
OmniEraser
28
6.832
21.3×
5.06
FLAIR
65
11.464
35.7×
7.83
Table 2. Inference efficiency and model size. SERO achieves the fastest runtime with single-step inference.
Figure 6. User study preference distribution across object-removal methods. A pie chart showing that our method, SERO, is highly preferred by users compared to existing object-removal baselines.
Figure 7. Qualitative effect of the semantic prior. Compared with the baseline, SERO produces more contextually coherent completions by conditioning bridge matching on the predicted hidden-region structure. Object-removal examples comparing the input, visible context, a baseline without a semantic prior, and SERO with the semantic prior.
Semantic Conditioning
FID ↓
localFID ↓
fMSE ↓
PSNR ↑
SSIM (%) ↑
Base (no semantic prior)
33.03
43.23
966.39
21.47
67.32
Base + CLIP features
29.45
40.19
867.64
21.85
67.49
Base + DINOv2 features
29.25
40.14
878.38
21.84
67.42
Base + I-JEPA predictive prior (Ours)
28.02
37.95
807.38
22.13
67.66
Table 3. Semantic prior ablation. I-JEPA outperforms CLIP and DINOv2 conditioning across all metrics.
Number of Steps
FID ↓
localFID ↓
fMSE ↓
PSNR ↑
SSIM(%) ↑
1 (ours)
28.02
37.95
807.38
22.13
67.66
2
28.67
37.92
853.08
21.94
67.25
3
40.34
58.28
929.65
21.92
67.11
4
31.25
40.63
924.70
21.69
66.78
Table 4. Effect of inference steps. One-step inference provides the best accuracy–efficiency trade-off, with additional steps yielding no consistent gains.
Real-world object removal is challenging due to two key difficulties: the target object's non-local effects, such as shadows and reflections, which are difficult to model, and the fact that user-provided masks are often inaccurate or incomplete. With billions of parameters and tens of denoising steps, diffusion-based models achieve strong removal performance at the expense of substantial computational cost, limiting their use in interactive applications and on edge devices. To address these challenges, we present OSOR (One-Step Object Removal), which simultaneously achieves efficient, effect-aware, and mask-robust object removal. Concretely, OSOR introduces: (1) an occupancy-guided discriminator for precise boundary supervision, enabling stable single-step diffusion training; (2) an alpha head that leverages knowledge from pretrained diffusion models to predict appropriate removal regions with minimal overhead, thereby handling imperfect masks; and (3) a semantic-anchored verification pipeline (SAVP) that filters noisy instruction-based triplets to produce effect-aware supervision at scale. Using SAVP, we curate CORNE, which contains 280K verified removal pairs, and further annotate AnimeEraseBench and TextEraseBench to evaluate performance on more complex removal tasks. Experiments show that OSOR surpasses strong multi-step diffusion baselines in perceptual quality while achieving 4× to 30× faster inference.
Qinming Zhou, Chenxi Sun, Deyang Kong +6
Shanghai Jiao Tong University · Tsinghua University · University of Electronic Science and Technology of China +3
Removing an object from a real image requires more than synthesizing plausible content within a mask: the method must suppress residual object features, preserve the unedited scene, and generate replacement content that is consistent with the surrounding background. This paper approaches object removal from a stage-based perspective and proposes a zero-shot framework for constrained latent inpainting with a frozen pretrained Stable Diffusion model, requiring no task-specific training or model fine-tuning. The method integrates SAM-based mask construction, BLIP image-caption conditioning, DDIM inversion, background-weighted masked null-text optimization, decoder self-attention masking, hard outside-mask latent anchoring, and localized renoise--denoise refinement into a unified pipeline. The method is evaluated through qualitative examples, quantitative local-consistency metrics, and ablation studies. The results demonstrate effective object removal and context-consistent replacement content. The ablations indicate that background-weighted masked NTI is particularly beneficial for structurally complex backgrounds, whereas the no-NTI variant is sufficient in other evaluated examples. Repeated refinement further reduces object remnants and boundary artifacts remaining after the primary editing pass.
Existing object removal tools often rely on manual masks or text prompts, making precise removal difficult for non-expert users in complex scenes and often leading to incomplete removal or unnatural background completion. To address this issue, we present ClickRemoval, an open-source interactive object removal tool built on pretrained Stable Diffusion models and driven solely by user clicks. Without additional training, hand-drawn masks, or text descriptions, ClickRemoval localizes target objects and restores the background through self-attention modulation during denoising. Experiments show that ClickRemoval achieves competitive results across quantitative metrics and user studies. We release a complete software package at https://github.com/zld-make/ClickRemoval under the Apache-2.0 license.
Ledun Zhang, Yatu Ji, Xufei Zhuang +1
Inner Mongolia University of Technology Hohhot, Inner Mongolia, China