Pith. sign in

REVIEW 4 major objections 4 minor 66 references

DORS: Dynamic Attention Routing for Diffusion-based Object Removal in Dense Scenes

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read DORS shows that pruning self-attention paths to look-alike instances fixes object removal in crowded scenes.

desk verdict The idea is sensible, but the published routing equations cap the filtered branch below 50% and cannot produce the claimed tenfold MSN drop — the central mechanism needs a clear fix before I'd trust the numbers. read the letter →

arxiv 2607.16656 v1 pith:FB3Y7TNW submitted 2026-07-18 cs.CV

Haitong Tang , Haipeng Liu , Yang Wang This is my paper
classification cs.CV
keywords objectremovaldensescenesdiffusionmodelsself-attentioninstanceinterferenceattentionroutinginpaintingtraining-free
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that object removal in dense scenes fails because diffusion self-attention lets masked query tokens align with visually similar instances outside the mask, leaking target semantics back into the hole. It proposes DORS, a training-free plug-in that prunes those attention connections and then selectively restores context through a spatially adaptive routing rule. On a new dense-scene benchmark and a public benchmark, the method reports large reductions in incomplete removals and duplicate artifacts, with removal-error metrics near zero. The implication is that dense-scene object removal can be treated as an information-flow control problem rather than a retraining problem.

What carries the argument

The load-bearing mechanism is a pair of attention biases. IFA builds a pruned bias matrix B_p that sets to a large negative value the attention from masked queries to both the masked region and the similar-instance mask M_s, and CGR computes a fused output O = w(x) O_p + (1-w(x)) O_f, where the per-pixel weight w(x) is a piecewise-linear function of distance to each similar instance, scaled by a global density ratio rho. This turns attention control into a spatially varying routing decision.

What would settle it

Measure the attention weights of masked queries in a diffusion model before any routing is applied, on a dense scene with known similar instances. If masked queries do not disproportionately attend to the similar-instance tokens compared to random background tokens, then instance interference is not the dominant failure cause and the pruning mechanism is not doing the work.

Watch

Extended reading notes

Core claim

The central claim is that instance interference in self-attention is the main cause of failed object removal in dense scenes, and that two complementary inference-time modifications eliminate it. Instance-Filtered Attention (IFA) adds a mask-guided bias to self-attention so that queries in the removal region cannot attend to tokens inside the target mask or to similar-instance masks supplied by a segmentation prompt. Context-Guided Routing (CGR) fuses this filtered branch with a full-context branch using a per-location weight that grows with global density of similar instances and with distance from them, preserving structural context near the mask while suppressing misleading semantics fart

Load-bearing premise

The method assumes the segmentation model, prompted with the target mask, correctly identifies every similar instance whose appearance actually draws harmful attention; if that mask misses a distractor, the pruned attention leaves an open channel for interference.

Editorial extensions

If this is right

  • Training-free object removal can match or surpass fine-tuned removal models on dense scenes, so expensive per-task training is not required for this failure mode.
  • The improvement transfers across diffusion backbones (SD1.5, SD2.0, SDXL) as a plug-in module, suggesting the attention-routing mechanism is backbone-agnostic.
  • The new DOR-Bench provides a stress test for instance interference, giving the community a benchmark where generic inpainting methods are known to fail.
  • Because control is applied only in early denoising steps, the method adds modest overhead (roughly 1.4s per image over the base model) while keeping later detail refinement untouched.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same shield-and-bridge routing could generalize to other attention-based image editing tasks, such as object insertion or swap, where semantic leakage from similar regions causes artifacts.
  • The dependency on a segmentation model to supply similar-instance masks is the clearest bottleneck; replacing that hand-off with a learned attention-attractor detector could make the method end-to-end and reduce the gap between oracle masks and practical ones.
  • A testable extension is to learn the hyperparameters alpha, beta1, beta2 from data instead of fixing them, which could improve robustness to scenes with extreme density or scale variations.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes DORS, a training-free, plug-and-play method for object removal in dense scenes. It introduces two attention-control components: Instance-Filtered Attention (IFA), which prunes self-attention from masked queries to similar-instance regions using SAM-derived masks, and Context-Guided Routing (CGR), which fuses the filtered branch with a full-context branch via a spatially adaptive weight. The authors also construct a new benchmark, DOR-Bench, consisting of 400 image-mask pairs with dense similar instances, and report state-of-the-art removal accuracy and background fidelity results on both DOR-Bench and the public RORD benchmark.

Significance. If the mechanism works as described, DORS addresses a genuinely important failure mode in object removal: instance interference caused by visually similar surroundings. The paper contributes a plausible attention-routing formulation, a new benchmark, and extensive experiments including plug-and-play evaluation across backbones and robustness to different SAM variants. The external RORD results and the consistent improvements across backbone variants provide some independent support beyond the author-created DOR-Bench. However, the central claim is weakened by an internal inconsistency in the routing equations and by evaluation choices that make the headline numbers less persuasive than they initially appear.

major comments (4)
  1. [§2.3, Eq. (10)–(15)] The published routing equations are internally under-powered relative to the reported results. Eq. (10) defines rho = area(M_s)/area(unmasked region), so rho <= 1; Eq. (11) sets wmax = alpha*rho, and the headline experiments use alpha = 0.5 (Table 10). Hence wmax <= 0.5 at every pixel. Eq. (15) then computes O = w(x)*O_p + (1-w(x))*O_f, meaning the filtered branch O_p is never the majority component. Table 5 reports FCB (O_f alone) at MSN 33.50 / MARS 24.83, IFA (O_p alone) at 11.75 / 5.43, and the full DORS at 1.25 / 0.70. A convex combination that is at least 50% O_f cannot plausibly produce errors dramatically lower than both components without some additional nonlinearity or an effective w much larger than Eq. (11) permits. The paper does not disclose any renormalization or alternative rho definition. This is a load-bearing inconsistency: either the equations need correction, or the
  2. [Table 10 vs. Table 1; Sec. H] The density-control hyperparameter alpha is selected by sweeping on DOR-Bench (Table 10) and the resulting best value (alpha = 0.5) is then used to produce the headline DOR-Bench numbers in Table 1. This is a form of evaluation leakage on the main benchmark. The paper should either report performance on a separate validation split, provide a sensitivity analysis that does not select alpha on the test set, or clearly state that DOR-Bench results are optimistic relative to a setting where alpha is fixed a priori. The external RORD results help, but they are not sufficient to rescue the DOR-Bench headline.
  3. [Tables 1–3 and §3.2] No error bars or repeated-run statistics are provided for any of the quantitative comparisons. Diffusion-based sampling is stochastic, and the reported metrics (MSN, MARS, PSNR, etc.) are known to vary across seeds. Without repeated runs or at least standard deviations, it is impossible to judge whether the large differences on DOR-Bench (e.g., MSN 1.25 vs. 6.50 for the second-best method) are significant or could overlap under noise. This is especially important because the method is training-free and uses a stochastic reverse process.
  4. [§2.2, Eq. (8); Table 11] The load-bearing assumption is that the SAM-based similar-instance mask M_s correctly identifies the regions whose visual appearance drives harmful attention. The robustness study (Table 11) measures instance-level mask quality (Recall/Precision/IoU) and its downstream effect on removal metrics, but it does not measure the correspondence between M_s and the actual attention-attracting regions. If M_s is incomplete or over-inclusive, the pruning in Eq. (8) may leave interference channels open or suppress useful context, yet the paper does not directly validate that the SAM masks align with the attention pathway that causes the artifacts. A simple attention-map correlation analysis or an experiment with deliberately corrupted masks would strengthen this assumption.
minor comments (4)
  1. [Throughout] The manuscript contains several typos and stylistic issues, e.g., 'reinSec. I' (Appendix), 'evalution' (Sec. D), and inconsistent use of 'DORS' vs. 'DAR' in captions. These should be corrected in a revision.
  2. [Fig. 1 and Fig. 5 captions] The qualitative figures are dense and the green bounding boxes are hard to see at small scale. Please ensure the figures are legible in the final version, possibly with higher-resolution crops or zoom insets.
  3. [Algorithm 1] The algorithm uses 'Step' for the update rule but the paper does not specify which solver is used (DDIM, DPM-Solver, etc.). The 20-step setting is mentioned, but a precise solver reference would improve reproducibility.
  4. [References] A few references are incomplete or informal, e.g., the Object Removal LoRA of Flux Fill Dev is cited with a GitHub user name and no venue, and FLUX.1-Fill-dev is cited with only a homepage. These should be completed or replaced with peer-reviewed sources where possible.

Circularity Check

0 steps flagged · score 1.0 of 10

No material circularity: DORS is an attention-engineering contribution; its central effects are explicitly constructed and independently validated on RORD.

full rationale

The paper does not present a formal derivation from first principles; it proposes an inference-time attention manipulation. The claimed mechanism is directly implemented in Eq. (8) (pruning attention to similar-instance masks) and Eq. (15) (convex fusion of filtered and full-context branches). The claim that IFA 'effectively eliminates misleading semantic references' is a description of the bias construction rather than a derived prediction, so it cannot reduce to its inputs in the circularity sense. The component ablation (Table 5) isolates FCB, IFA, and CGR, and the external RORD benchmark (Table 2) is not constructed by the authors and uses the same fixed settings; this provides independent support for the central claim. The only evaluative caveat is that the density-control parameter alpha is selected on DOR-Bench (Sec. H, Table 10) and the headline DOR-Bench numbers use that choice; this is a hyperparameter-selection/benchmark-contamination concern rather than a circular derivation, and it is mitigated by the RORD results. The skeptic's concern that Eq. (11) caps wmax at most 0.5 and hence Eq. (15) cannot produce the reported large gains is an internal-consistency/implementation-disclosure concern, not a circularity: the equation as written may be under-powered, but that is a correctness risk, not equivalence of output to input. No self-citation chain is load-bearing: prior work is cited for standard components, and the novel contribution is the specific mask-guided pruning and routing.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method depends on three load-bearing choices: (1) the equivalence between visual instance similarity and harmful attention attraction, (2) SAM's ability to localize those instances from a mask prompt, and (3) the schedule and strength constants of the routing. The first is the scientific premise; the second is tooling; the third injects hand-tuned parameters, one of which (alpha) is tuned on the evaluation benchmark.

free parameters (4)
  • alpha (density control) = 0.5
    Eq. (11) sets maximum filtering strength as alpha*rho; Table 10 sensitivity analysis on DOR-Bench picks alpha=0.5, which is then used for final DOR-Bench results in Table 1.
  • beta1, beta2 (piecewise routing thresholds) = not reported
    Eq. (13) introduces 0<beta1<beta2<1 to control transition distances; no values or sensitivity analysis are given in the provided text, yet they determine the routing weight.
  • stage-wise intervention cutoff (T/2) = first half of 20 steps (2/4 schedule)
    Algorithm 1 applies DAR only for t>T/2; Table 9 ablation on schedules selects 2/4 steps as best on DOR-Bench.
  • attention bias constant tau = unspecified 'large negative constant'
    Eqs. (6) and (8) use tau to suppress attention; the exact value is not stated in the paper.
assumptions (4)
  • domain assumption Self-attention similarity matching in the pre-trained diffusion model is the mechanism that transfers target semantics into masked regions; suppressing attention to visually similar instances removes that transfer.
    Central premise in Sec 2.1-2.2; motivated by attention visualizations (Fig. 3), but not formally established.
  • ad hoc to paper SAM prompted with the target mask segments all and only the instances whose visual similarity causes interference (Ms).
    Section 2.2 Eq (8) and Fig. 4 rely on Ms from SAM3; Table 11 shows performance degrades with weaker segmenters, so the method's gain is coupled to segmentation quality.
  • domain assumption Early denoising steps establish global structure and later steps refine details, so attention intervention can be limited to the first half of the schedule.
    Section 2.3 Stage-wise Application and Algorithm 1; supported by diffusion folklore and Table 9 ablations, not by a separate analysis.
  • domain assumption DOR-Bench samples are representative of the dense-scene object removal problem and metric scores are comparable across methods when run with official implementations.
    Appendix B describes curation by the authors: 400 pairs selected from public datasets [14,21] and manually refined; no independent release or third-party evaluation is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DORS: Dynamic Attention Routing for Diffusion-based Object Removal in Dense Scenes." pith.science (2026). https://pith.science/paper/FB3Y7TNW

@misc{pith2026260716656,
  author       = {Pith},
  title        = {Pith review of: DORS: Dynamic Attention Routing for Diffusion-based Object Removal in Dense Scenes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FB3Y7TNW}},
  note         = {Machine review of arXiv:2607.16656}
}
read the original abstract

Object removal aims to eliminate target objects specified by a mask while preserving visual consistency with the surrounding regions. Existing methods typically rely on contextual information from surrounding regions. However, in dense scenes where the surrounding regions contain instances visually similar to the removal target, such reliance often leads to semantic interference, resulting in incomplete removal. This problem arises from erroneous information propagation in the attention space, where masked queries tend to align with such instances due to global similarity matching in self-attention. To address this challenge, we propose a Diffusion-based Object Removal framework for dense Scenes, dubbed DORS, built upon a Dynamic Attention Routing mechanism comprising two complementary components: Instance-Filtered Attention (IFA), which suppresses misleading semantic information from similar instances through dynamically constructed mask-guided attention constraints, and Context-Guided Routing (CGR), which dynamically routes complementary scene information to maintain visual consistency. We further introduce DOR-Bench, a benchmark tailored for object removal in dense scenes. Extensive experiments demonstrate that DORS outperforms state-of-the-art methods, particularly in reducing incomplete removal and duplicate artifacts. The code will be available at https://github.com/httang1224/DORS.

Figures

Figures reproduced from arXiv: 2607.16656 by the authors.

Figure 1
Figure 1. Object removal in dense scenes containing multiple similar instances remains challenging for state-of-the-art methods, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of DORS. It introduces a plug-and-play Dynamic Attention Routing (DAR) mechanism into the denoising [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Self-attention visualization and corresponding re [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Qualitative comparison of object removal results in DOR-Bench, our method (DORS) achieves more accurate target [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 4
Figure 4. Figure 4: This spatially adaptive weight controls the contribution of [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison of DORS with different [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Additional qualitative results under various challenging scenarios. Zoom in for better visualization of fine details. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: User study interface. The input image with the [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Additional qualitative ablation results. Mask-based [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Failure cases of DORS on complex visual effects. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 9 linked inside Pith

  1. [1]

    Omri Avrahami, Ohad Fried, and Dani Lischinski. 2023. Blended latent diffusion. ACM transactions on graphics (TOG)42, 4 (2023), 1–11

    2023

  2. [2]

    Omri Avrahami, Or Patashnik, Ohad Fried, Egor Nemchinov, Kfir Aberman, Dani Lischinski, and Daniel Cohen-Or. 2025. Stable flow: Vital layers for training-free image editing. InProceedings of the Computer Vision and Pattern Recognition Conference. 7877–7888

    2025

  3. [3]

    Christopher M. Bishop. 2006.Pattern Recognition and Machine Learning. Springer

    2006

  4. [4]

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, An- drew Huang, et al. 2025. Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719(2025)

    arXiv 2025

  5. [5]

    Aditya Chandrasekar, Goirik Chakrabarty, Jai Bardhan, Ramya Hebbalaguppe, and Prathosh AP. 2024. Remove: A reference-free metric for object erasure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 7901–7910

    2024

  6. [6]

    Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao

  7. [7]

    Zhekai Chen, Wen Wang, Zhen Yang, Zeqing Yuan, Hao Chen, and Chunhua Shen. 2024. Freecompose: Generic zero-shot image composition with diffusion prior. InEuropean Conference on Computer Vision. Springer, 70–87

    2024

  8. [8]

    Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis.Advances in neural information processing systems34 (2021), 8780–8794

    2021

Show all 66 references
  1. [9]

    Yiğit Ekin, Ahmet B Yildirim, Erdem E Caglar, Aykut Erdem, Erkut Erdem, and Aysegul Dundar. 2024. Clipaway: Harmonizing focused embeddings for removing objects via diffusion models.Advances in Neural Information Processing Systems 37 (2024), 17572–17601

    2024

  2. [10]

    Xinjian Gao, Yang Wang, and Meng Wang. 2023. Macroscopic-and-microscopic rain streaks disentanglement network for single-image deraining.IEEE Transac- tions on Image Processing32 (2023), 2663–2677

    2023

  3. [11]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems30 (2017)

    2017

  4. [12]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

    2020

  5. [13]

    Jonathan Ho and Tim Salimans. 2022. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598(2022)

    2022 arXiv

  6. [14]

    Longtao Jiang, Zhendong Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Lei Shi, Dong Chen, and Houqiang Li. 2025. Smarteraser: Remove anything from images using masked-region guidance. InProceedings of the Computer Vision and Pattern Recognition Conference. 24452–24462

    2025

  7. [15]

    Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. 2024. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. InEuropean Conference on Computer Vision. Springer, 150–168

    2024

  8. [16]

    Diederik P Kingma and Max Welling. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114(2013)

    2013 arXiv

  9. [17]

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al

  10. [18]

    Jari Korhonen and Junyong You. 2012. Peak signal-to-noise ratio revisited: Is simple beautiful?. In2012 Fourth international workshop on quality of multimedia experience. IEEE, 37–38

    2012

  11. [19]

    Black Forest Labs. 2024. FLUX. https://github.com/black-forest-labs/flux

    2024

  12. [20]

    Fan Li, Zixiao Zhang, Yi Huang, Jianzhuang Liu, Renjing Pei, Bin Shao, and Songcen Xu. 2024. Magiceraser: Erasing any objects via semantics-aware control. InEuropean Conference on Computer Vision. Springer, 215–231

    2024

  13. [21]

    Ruibin Li, Tao Yang, Song Guo, and Lei Zhang. 2025. Rorem: Training a robust object remover with human-in-the-loop. InProceedings of the Computer Vision and Pattern Recognition Conference. 14024–14035

    2025

  14. [22]

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le

  15. [23]

    Heng Liu, Guanghui Li, Mingqi Gao, Xiantong Zhen, Feng Zheng, and Yang Wang. 2025. Few-shot referring video single-and multi-object segmentation via cross-modal affinity with instance sequence matching.International Journal of Computer Vision133, 8 (2025), 5610–5628

    2025

  16. [24]

    Haipeng Liu, Yang Wang, Biao Qian, Meng Wang, and Yong Rui. 2024. Struc- ture matters: Tackling the semantic discrepancy in diffusion models for image inpainting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 8038–8047

    2024

  17. [25]

    Haipeng Liu, Yang Wang, and Meng Wang. 2025. One Stone with Two Birds: A Null-Text-Null Frequency-Aware Diffusion Models for Text-Guided Image Inpainting. InAdvances in Neural Information Processing Systems, Vol. 38. 10833– 10859

    2025

  18. [26]

    Haipeng Liu, Yang Wang, Meng Wang, and Yong Rui. 2022. Delving globally into texture and structure for image inpainting. InProceedings of the 30th ACM International Conference on Multimedia. 1270–1278

    2022

  19. [27]

    Xingchao Liu, Chengyue Gong, and Qiang Liu. 2022. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003(2022)

    2022 arXiv

  20. [28]

    Yi Liu, Hao Zhou, Benlei Cui, Wenxiang Shang, and Ran Lin. 2025. Erase dif- fusion: Empowering object removal through calibrating diffusion pathways. In Proceedings of the Computer Vision and Pattern Recognition Conference. 2418–2427

    2025

  21. [29]

    lrzjason. 2025. Object Removal LoRA of Flux Fill Dev

    2025

  22. [30]

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. 2025. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. Machine Intelligence Research22, 4 (2025), 730–751

    2025

  23. [31]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952 (2023)

    2023 arXiv

  24. [32]

    Biao Qian, Yang Wang, Richang Hong, and Meng Wang. 2023. Adaptive data-free quantization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 7960–7968

    2023

  25. [33]

    Biao Qian, Yang Wang, Richang Hong, and Meng Wang. 2023. Rethinking data- free quantization as a zero-sum game. InProceedings of the AAAI conference on artificial intelligence, Vol. 37. 9489–9497

    2023

  26. [34]

    Biao Qian, Yang Wang, Yong Wu, and Jungong Han. 2026. Selective Coupling of Decoupled Informative Regions: Masked Attention Alignment for Data-Free Quantization of Vision Transformers.arXiv preprint arXiv:2606.04373(2026)

    2026 arXiv

  27. [35]

    Biao Qian, Yang Wang, Hongzhi Yin, Richang Hong, and Meng Wang. 2022. Switchable online knowledge distillation. InEuropean Conference on Computer Vision. Springer, 449–466

    2022

  28. [36]

    Qwen Team. 2026. Qwen3.5: Towards Native Multimodal Agents. https://qwen. ai/blog?id=qwen3.5

    2026

  29. [37]

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al

  30. [38]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis With Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 10684–10695

    2022

  31. [39]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical image computing and computer-assisted intervention. Springer, 234–241

    2015

  32. [40]

    Min-Cheol Sagong, Yoon-Jae Yeo, Seung-Won Jung, and Sung-Jea Ko. 2022. RORD: A Real-world Object Removal Dataset.. InBMVC. 542

    2022

  33. [41]

    InInternational Conference on Learning Representations, Vol

    Sam 2: Segment anything in images and videos. InInternational Conference on Learning Representations, Vol. 2025. 28085–28128

    2025

  34. [42]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2020. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502(2020)

    2020 arXiv

  35. [43]

    Wenhao Sun, Xue-Mei Dong, Benlei Cui, and Jingqun Tang. 2025. Attentive eraser: Unleashing diffusion model’s object removal potential via self-attention redirection guidance. InProceedings of the AAAI Conference on Artificial Intelli- gence, Vol. 39. 20734–20742

    2025

  36. [44]

    Haitong Tang, Shuang He, Mengduo Yang, Xia Lu, Qin Yu, Kaiyue Liu, Hongjie Yan, and Nizhuan Wang. 2024. CSC-Unet: a novel convolutional sparse coding strategy based neural network for semantic segmentation.IEEE Access12 (2024), 35844–35854

    2024

  37. [45]

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. 2022. Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in neural in...

    2022

  38. [46]

    Yikai Wang, Chenjie Cao, Junqiu Yu, Ke Fan, Xiangyang Xue, and Yanwei Fu. 2025. Towards enhanced image inpainting: Mitigating unwanted object insertion and preserving color consistency. InProceedings of the Computer Vision and Pattern Recognition Conference. 23237–23248

    2025

  39. [47]

    Yang Wang, Jinjia Peng, Huibing Wang, and Meng Wang. 2022. Progressive learn- ing with multi-scale attention network for cross-domain vehicle re-identification. Science China Information Sciences65, 6 (2022), 160103

    2022

  40. [48]

    Yang Wang, Biao Qian, Haipeng Liu, Yong Rui, and Meng Wang. 2024. Unpacking the gap box against data-free knowledge distillation.IEEE Transactions on Pattern Analysis and Machine Intelligence46, 9 (2024), 6280–6291

    2024

  41. [49]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  42. [50]

    Runpu Wei, Zijin Yin, Shuo Zhang, Lanxiang Zhou, Xueyi Wang, Chao Ban, Tianwei Cao, Hao Sun, Zhongjiang He, Kongming Liang, et al. 2025. Omnieraser: Remove objects and their effects in images with paired video-frame data.arXiv preprint arXiv:2501.07397(2025)

    2025 arXiv

  43. [51]

    Daniel Winter, Matan Cohen, Shlomi Fruchter, Yael Pritch, Alex Rav-Acha, and Yedid Hoshen. 2024. Objectdrop: Bootstrapping counterfactuals for photoreal- istic object removal and insertion. InEuropean Conference on Computer Vision. Springer, 112–129

    2024

  44. [52]

    Liangbin Xie, Daniil Pakhomov, Zhonghao Wang, Zongze Wu, Ziyan Chen, Yuqian Zhou, Haitian Zheng, Zhifei Zhang, Zhe Lin, Jiantao Zhou, et al. 2025. Turbofill: adapting few-step text-to-image model for fast image inpainting. In Proceedings of the Computer Vision and Pattern Reco...

    2025

  45. [53]

    Navve Wasserman, Noam Rotstein, Roy Ganz, and Ron Kimmel. 2025. Paint by inpaint: Learning to add image objects by removing them first. InProceedings of the Computer Vision and Pattern Recognition Conference. 18313–18324. MM ’26, October 2026, Rio de Janeiro, Brazil Haitong Ta...

    2025

  46. [54]

    Yongsheng Yu, Ziyun Zeng, Haitian Zheng, and Jiebo Luo. 2025. Omnipaint: Mastering object-oriented editing via disentangled insertion-removal inpainting. InProceedings of the IEEE/CVF International Conference on Computer Vision. 17324– 17334

    2025

  47. [55]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

  48. [56]

    Xulu Zhang, Xiaoyong Wei, Wentao Hu, Jinlin Wu, Jiaxin Wu, Wengyu Zhang, Zhaoxiang Zhang, Zhen Lei, and Qing Li. 2025. A survey on personalized content synthesis with diffusion models.Machine Intelligence Research22, 5 (2025), 817– 848

    2025

  49. [57]

    Jinrui Yang, Qing Liu, Yijun Li, Soo Ye Kim, Daniil Pakhomov, Mengwei Ren, Jianming Zhang, Zhe Lin, Cihang Xie, and Yuyin Zhou. 2025. Generative image layer decomposition with visual effects. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. ...

    2025

  50. [58]

    Yixing Zhu, Qing Zhang, Yitong Wang, Yongwei Nie, and Wei-Shi Zheng. 2025. EntityErasure: Erasing Entity Cleanly via Amodal Entity Segmentation and Com- pletion. InProceedings of the Computer Vision and Pattern Recognition Conference. 28274–28283

    2025

  51. [59]

    Zixin Zhu, Haoxiang Li, Xuelu Feng, He Wu, Chunming Qiao, and Junsong Yuan

  52. [60]

    best_method

    Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. 2024. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. InEuropean Conference on Computer Vision. Springer, 195–211. DORS: Dynamic Attention Routing for Diffusion-...

    2024

  53. [62]

    Jixin Zhao, Shangchen Zhou, Zhouxia Wang, Peiqing Yang, and Chen Change Loy. 2025. Objectclear: Complete object removal via object-effect attention.arXiv preprint arXiv:2505.22636(2025)

    2025

  54. [65]

    GeoRemover: Removing Objects and Their Causal Visual Artifacts.arXiv preprint arXiv:2509.18538(2025)

    2025

  55. [2018]

    InProceedings of the IEEE conference on computer vision and pattern recognition

    The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition. 586–595

  56. [2022]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747 (2022)

    2022 arXiv

  57. [2023]

    InProceedings of the IEEE/CVF international conference on computer vision

    Segment anything. InProceedings of the IEEE/CVF international conference on computer vision. 4015–4026

  58. [2025]

    IEEE Transactions on Pattern Analysis and Machine Intelligence(2025)

    Anydoor: zero-shot image customization with region-to-region reference. IEEE Transactions on Pattern Analysis and Machine Intelligence(2025)

    2025

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.