Pith. sign in

REVIEW 5 major objections 4 minor 33 references

2D Instance Editing in 3D Space

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Point-drag 2D image editing that lifts the object to 3D before deforming preserves identity and consistency far better than pixel-space dragging.

desk verdict A plausible 2D-3D-2D editing pipeline that is under-specified exactly where it matters: the 2D-to-3D handle mapping and camera alignment are never defined, and the evaluation is purely anecdotal. read the letter →

arxiv 2507.05819 v1 pith:ST6MM6TW submitted 2025-07-08 cs.CV

Yuhuan Xie , Aoxuan Pan , Ming-Xian Lin , Wei Huang , Yi-Hua Huang , Xiaojuan Qi This is my paper
classification cs.CV
keywords imageediting3DGaussiansplattingas-rigid-as-possibledeformationpoint-baseddragsingle-imagegenerationinpaintingobjectidentitypreservation
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

This paper argues that the right way to edit a 2D image by dragging points is to leave the pixel plane entirely: segment the target object, lift it into a 3D Gaussian-splatting model, deform the 3D model under a local-rigidity constraint, and then re-render and inpaint it back into the original photo. The authors claim this 2D-3D-2D route makes large pose changes, rigid rotations, and non-rigid deformations physically plausible while preserving object identity, which they say DragGAN and DragDiffusion fail to do because they manipulate pixels in the image plane. If that claim holds, interactive image editing no longer needs category-specific generative models to stay consistent. The demonstration is qualitative, focusing on elephants, horses, vehicles, and man-made objects, with the paper's own limitation that everything depends on the quality of the single-image 3D reconstruction.

What carries the argument

The load-bearing mechanism is the 2D-3D-2D detour through an editable 3D Gaussian Splatting (3DGS) model, where as-rigid-as-possible (ARAP) deformation is applied to a sparse graph of control points rather than to millions of Gaussians. Farthest-point sampling picks 512 control points; connecting each to K=8 neighbors through shortest-path graph distances instead of raw Euclidean distance preserves topology across separate object regions. The ARAP energy penalizes any local deviation from rigid motion, and is minimized by alternating between solving a linear Laplacian system for the deformed positions and estimating local rotations by SVD; three iterations suffice. Linear blend skinning then propagates the control-point motion to every Gaussian, updating both positions and quaternion rotations, so edits stay smooth and physically plausible. This sparse-graph-plus-skinning design is what buys real-time interaction on a representation that would otherwise be too dense to deform directly.

What would settle it

Run the pipeline on a set of images where the target object is partially occluded or small in the frame, apply identical drag handles, and measure object-identity similarity against DragDiffusion; the paper's central claim would be falsified if, on inputs where reconstruction is visibly poor, edits become less consistent than the 2D baseline rather than more consistent.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that lifting an object instance into editable 3D geometry before applying a point-based edit converts a fundamentally ill-posed 2D image-manipulation problem into a well-posed 3D deformation problem. Given a masked crop of an object, an off-the-shelf single-image 3D generator produces a 3D Gaussian Splatting model; the method then samples 512 control points, builds a topology-aware graph, and minimizes an as-rigid-as-possible energy subject to user handle constraints, propagating the resulting deformation to all Gaussians via linear blend skinning. The edited 3D object is rendered from the original viewpoint and composited over an inpainted background with a second inpainting pass to remove seams. The authors' claim is that this procedure consistently handles large-scale edits across a wide variety of images, producing high-fidelity outputs that robustly preserve object identity and scene consistency, outperforming DragGAN and DragDiffusion on pose changes and non-rigid deformations.

Load-bearing premise

Everything downstream—deformation, re-rendering, inpainting, identity preservation—assumes the single-image 3D reconstruction of the masked object is accurate enough to edit; occlusions, harsh lighting, or distant objects that degrade reconstruction degrade the entire result.

Editorial extensions

If this is right

  • Large viewpoint and pose changes, which tend to break pixel-space drag methods, become natural operations once the object is represented as a rigidly deformable 3D model.
  • The same pipeline works on arbitrary segmented objects, so interactive editing no longer requires category-specific generative models.
  • The sparse control-point graph keeps the deformation solve fast enough for real-time interaction, letting users refine edits iteratively.
  • Because the object is positioned with an explicit 6DoF pose, moving an object within the scene is supported in the same pass as reshaping it.
  • Any improvement in single-image 3D generation or inpainting translates directly into better edit fidelity, since the editing stage itself is a generic constrained deformation.

Reading between the lines

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

  • Beyond the paper's argument, the same 2D-3D-2D pattern should extend to video editing: lift the object once, deform it once, and re-render over frames, which would give temporally consistent drag-style edits that pixel-space methods lack.
  • A quantitative test of the core claim is within reach: a standardized drag benchmark with identity-similarity metrics would let one plot edit fidelity against reconstruction quality, and the paper's argument predicts a sharp drop exactly where lifting degrades.
  • The modular design implies that the editing stage is generic: any future single-image 3D generator or inpainting model that outperforms the current components would automatically improve this pipeline without redesigning the deformation.
  • On out-of-distribution objects where the single-image 3D prior is weak, users may in practice prefer the very 2D baselines the paper seeks to replace, so the practical win is conditional on reconstruction quality.
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

5 major / 4 minor

Summary. The paper proposes a 2D-3D-2D image editing pipeline: the target object is segmented with SAM, lifted to a 3D Gaussian Splatting representation by TRELLIS, deformed under an as-rigid-as-possible (ARAP) energy with sparse control points and linear blend skinning, and then rendered back into the original image with inpainting to fill the background and blend boundaries. The authors claim that this approach outperforms DragGAN and DragDiffusion on large pose changes and non-rigid deformations while preserving object identity and scene consistency, and that it enables real-time interactive editing. The ARAP derivation in Eqs. (1)-(5) is standard and coherent, but the comparative and real-time claims are supported only by a few qualitative examples, and the interface between 2D user input and the 3D deformation is not specified.

Significance. The central idea of editing an object in a reconstructed 3D representation and then compositing it back into the original 2D image is a plausible and potentially useful alternative to pixel-space drag editing. The paper also contains a reasonable topology-aware graph construction (Sec. 3.2, Fig. 3) and integrates off-the-shelf components such as SAM, TRELLIS, and inpainting. However, the significance is not currently demonstrated: there are no quantitative metrics, no user study, no timing measurements, and no ablations, and the missing 2D-to-3D mapping and camera-alignment procedure make the method non-reproducible as described. These issues are load-bearing because the paper's main claims are superiority over prior methods and real-time interaction.

major comments (5)
  1. [Section 3.2, Eq. (2)] The user interaction is described as clicking on the image and dragging control points 'to reposition them in 3D space,' but the paper never specifies how 2D pixel coordinates are unprojected onto 3D Gaussian control points, nor how a 2D drag vector is converted into a 3D translation, particularly along the depth axis. As written, Eq. (2) is not instantiable from the stated input, so the central capability of point-guided 3D editing cannot be reproduced or evaluated.
  2. [Section 3.3] The method states that the edited object is rendered 'from a consistent viewpoint' and placed at a 'specified 6DoF pose,' but no procedure is given for estimating the camera pose, scale, rotation, and translation that align TRELLIS's canonical reconstruction with the original image. TRELLIS normalizes object pose and scale, so an unknown rigid transform remains; without specifying how this transform is obtained, the claimed scene consistency is not well-defined and the pipeline cannot be reproduced.
  3. [Section 4.1] The central claim that the method 'significantly outperforms' DragGAN and DragDiffusion is supported only by qualitative examples in Figs. 4 and 5. There are no quantitative metrics, error bars, user studies, or statistical tests. The paper needs at least control-point accuracy, identity/consistency measures (e.g., LPIPS or a user study), and a discussion of failure cases to substantiate the comparative claim.
  4. [Sections 1 and 4] The paper repeatedly claims 'real-time' interactive editing, but it reports no runtime measurements, hardware configuration, or interaction latency. Since real-time operation is stated as a contribution, this claim is currently unsupported.
  5. [Eq. (5)] The blended quaternion q'_i is a weighted sum of quaternions, which is not generally unit-norm and therefore does not always represent a valid rotation. Without an explicit normalization step or a different rotation-blending scheme, the deformed Gaussian orientations are not well-defined, which undermines the claimed physical plausibility of the deformation.
minor comments (4)
  1. [Section 3.2] The phrase 'K^2-nearest neighbors' with K=8 is ambiguous because 'K^2' could be read as the square of the scalar K (64) or as a separate parameter; please clarify the notation.
  2. [References] References [22] and [23] are duplicate entries for DragGAN with inconsistent author lists and conference details; please consolidate them into a single reference.
  3. [Section 3.3] The phrase 'The drop stage' appears to be a typo; it should likely be 'the placement stage' or 'the second stage' for clarity.
  4. [Figure 5] The caption states that DragGAN failed to produce any results for the man-made object images; please clarify whether this is due to the official implementation, model limitations, or a configuration choice, since this affects the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the 2D-3D-2D pipeline is composed of external lifting and inpainting tools plus a standard ARAP deformation, so the central claims are not equivalent to the method's inputs.

full rationale

The paper's derivation chain is self-contained rather than circular. The 2D-to-3D step uses SAM and TRELLIS as off-the-shelf external systems, and the editing step minimizes the standard ARAP rigidity energy (Eq. 1) under user constraints (Eq. 2), solved by alternating SVD rotation estimation and a linear Laplacian solve (Eqs. 3-4), followed by LBS skinning (Eq. 5). None of these equations is defined in terms of the claimed output (identity-preserving large-scale edits), and no parameter is fitted to the qualitative evaluation images. The one self-citation relevant to the deformation, SC-GS [11], is a prior published work by two of the authors, but the energy is also attributed to the independent ARAP literature [26], and the cited method does not smuggle in the present paper's 2D editing claims; it is reused as a component. Missing details such as how 2D clicks are unprojected to 3D handle positions and how TRELLIS's canonical frame is aligned to the original camera are specification gaps that affect reproducibility, and the conclusion's stated dependence on reconstruction quality is an acknowledged limitation, but neither constitutes circularity. Therefore no step reduces by construction to its own input.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on two external black boxes: TRELLIS for single-image 3D generation and off-the-shelf inpainting models for background and seam filling. The deformation itself is standard ARAP optimization with hand-set hyperparameters. There are no newly invented entities or fitted physical constants.

free parameters (4)
  • control point count = 512
    Chosen by hand in Sec 3.2; determines deformation granularity and real-time cost.
  • rigidity graph neighborhood size K = 8
    K=8 nearest neighbors per control point in Eq. 1, chosen in Sec 3.2.
  • LBS neighborhood size K~ = 3
    Three nearest control points used to skin each Gaussian in Eq. 5, chosen in Sec 3.2.
  • ARAP iterations = 3
    Number of alternating solve and rotation-update iterations, stated in Sec 3.2.
assumptions (5)
  • domain assumption Local rigidity prior for deformation
    Sec 3.2 assumes deformation preserves local rigidity; plausible for many objects but not for fluid-like or loosely articulated shapes.
  • domain assumption TRELLIS produces a faithful 3DGS from a single view
    Sec 3.1 relies on off-the-shelf TRELLIS; reconstruction quality is load-bearing and acknowledged as a limitation in Sec 5.
  • domain assumption SAM segmentation is accurate enough
    Sec 3.1 uses SAM masks to define the object and the background hole; segmentation errors propagate through the pipeline.
  • domain assumption Inpainting models restore background and seams consistently
    Sec 3.3 delegates background and seam inpainting to off-the-shelf models; failures are acknowledged in Sec 5.
  • domain assumption Control-point graph preserves topology via shortest-path K-NN
    Sec 3.2 builds a graph on K/2 nearest neighbors then shortest paths; this prevents cross-region links but can still fail on thin or intricate structures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of 2D Instance Editing in 3D Space." pith.science (2026). https://pith.science/paper/ST6MM6TW

@misc{pith2026250705819,
  author       = {Pith},
  title        = {Pith review of: 2D Instance Editing in 3D Space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ST6MM6TW}},
  note         = {Machine review of arXiv:2507.05819}
}
read the original abstract

Generative models have achieved significant progress in advancing 2D image editing, demonstrating exceptional precision and realism. However, they often struggle with consistency and object identity preservation due to their inherent pixel-manipulation nature. To address this limitation, we introduce a novel "2D-3D-2D" framework. Our approach begins by lifting 2D objects into 3D representation, enabling edits within a physically plausible, rigidity-constrained 3D environment. The edited 3D objects are then reprojected and seamlessly inpainted back into the original 2D image. In contrast to existing 2D editing methods, such as DragGAN and DragDiffusion, our method directly manipulates objects in a 3D environment. Extensive experiments highlight that our framework surpasses previous methods in general performance, delivering highly consistent edits while robustly preserving object identity.

Figures

Figures reproduced from arXiv: 2507.05819 by the authors.

Figure 1
Figure 1. Given 2D instances in images, we propose a method to lift them into 3D, enabling edits under the local rigidity assumption. This [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Our method involves lifting 2D instances into 3D space, editing them under the assumption of local rigidity, and then repositioning [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Graph better preserves topology than the Euclidean. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative comparisons of our method with DragDiffusion [ [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparisons of our method with DragDiffusion [ [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Diverse editing results produced by our method. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 27 canonical work pages

  1. [1]

    Mpb: A mod- ified poisson blending technique

    Mahmoud Afifi and Khaled F Hussain. Mpb: A mod- ified poisson blending technique. Computational Vi- sual Media, 1:331–341, 2015. 2

    work page 2015

  2. [2]

    Image inpainting with the navier-stokes equations

    Wilson Au and Ryo Takei. Image inpainting with the navier-stokes equations. Available at Final Report APMA, 930, 2001. 2

    work page 2001

  3. [3]

    Seam carving for content-aware image resizing

    Shai Avidan and Ariel Shamir. Seam carving for content-aware image resizing. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2 , pages 609–617. 2023. 2

    work page 2023

  4. [4]

    Instructpix2pix: Learning to follow image edit- ing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image edit- ing instructions. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 18392–18402, 2023. 1, 2

    work page 2023

  5. [5]

    Efficient geometry-aware 3d generative adversarial networks

    Eric R Chan, Connor Z Lin, Matthew A Chan, Matthew Tancik, Eric Ng, Tero Aumentado- Armstrong, and Avinash Kulkarni. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 16123–16133, 2022. 3

    work page 2022

  6. [6]

    3d-r2n2: A unified approach for single and multi-view 3d object recon- struction

    Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object recon- struction. In Computer vision–ECCV 2016: 14th Eu- ropean conference, amsterdam, the netherlands, Oc- 6 tober 11-14, 2016, proceedings, part VIII 14 , pages 628–644. Springer, 2016. 2

    work page 2016

  7. [7]

    Cyclegan, a master of steganography

    Casey Chu, Andrey Zhmoginov, and Mark Sandler. Cyclegan, a master of steganography. arXiv preprint arXiv:1712.02950, 2017. 2

    arXiv 2017

  8. [8]

    Instruct- nerf2nerf: Editing 3d scenes with instructions

    Ayaan Haque, Matthew Tancik, Alexei A Efros, Alek- sander Holynski, and Angjoo Kanazawa. Instruct- nerf2nerf: Editing 3d scenes with instructions. In Pro- ceedings of the IEEE/CVF International Conference on Computer Vision, pages 19740–19750, 2023. 3

    work page 2023

Show all 33 references
  1. [9]

    Pix2pix gan for image-to-image translation

    Joyce Henry, Terry Natalie, and Den Madsen. Pix2pix gan for image-to-image translation. Research Gate Publication, pages 1–5, 2021. 2

    2021

  2. [10]

    Prompt- to-prompt image editing with cross attention control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aber- man, Yael Pritch, and Daniel Cohen-Or. Prompt- to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2

    2022 arXiv

  3. [11]

    Sc-gs: Sparse- controlled gaussian splatting for editable dynamic scenes

    Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse- controlled gaussian splatting for editable dynamic scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 4220–4230, 2024. 3

    2024

  4. [12]

    A style- based generator architecture for generative adversar- ial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style- based generator architecture for generative adversar- ial networks. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 4401–4410, 2019. 2

    2019

  5. [13]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139–1, 2023. 3

    2023

  6. [14]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 2023. 2, 3

    2023

  7. [15]

    Reference-based image composition with sketch via structure-aware diffusion model.arXiv preprint arXiv:2304.09748, 2023

    Kangyeol Kim, Sunghyun Park, Junsoo Lee, and Jaegul Choo. Reference-based image composition with sketch via structure-aware diffusion model.arXiv preprint arXiv:2304.09748, 2023. 1

    2023 arXiv

  8. [16]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vi- sion, pages 4015–4026, 2023. 2, 3

    2023

  9. [17]

    Zero-1-to-3: Zero- shot one image to 3d object

    Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, and Carl V ondrick. Zero-1-to-3: Zero- shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 2

    2023 arXiv

  10. [18]

    Joint-dependent local deformations for hand animation and object grasping

    Nadia Magnenat-Thalmann, Richard Laperri `ere, and Daniel Thalmann. Joint-dependent local deformations for hand animation and object grasping. In Proceed- ings on Graphics interface’88, pages 26–33, 1989. 4

    1989

  11. [19]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM , 65(1): 99–106, 2021. 2

    2021

  12. [20]

    The expectation-maximization algo- rithm

    Todd K Moon. The expectation-maximization algo- rithm. IEEE Signal processing magazine , 13(6):47– 60, 1996. 4

    1996

  13. [21]

    Dragondiffusion: Enabling drag- style manipulation on diffusion models.arXiv preprint arXiv:2307.02421, 2023

    Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Dragondiffusion: Enabling drag- style manipulation on diffusion models.arXiv preprint arXiv:2307.02421, 2023. 2

    2023 arXiv

  14. [22]

    Drag your gan: Interactive point- based manipulation on the generative image manifold

    Xingang Pan, Ayush Tewari, Thomas Leimk ¨uhler, Kripasindhu Aberman, Christian Theobalt, and Jaakko Lehtinen. Drag your gan: Interactive point- based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings ,

    2023

  15. [23]

    Drag your gan: Interactive point-based manipulation on the generative image manifold

    Xingang Pan, Ayush Tewari, Thomas Leimk ¨uhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIG- GRAPH 2023 conference proceedings , pages 1–11,

    2023

  16. [24]

    High- resolution image synthesis with latent diffusion mod- els

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High- resolution image synthesis with latent diffusion mod- els. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 10684–10695, 2022. 2

    2022

  17. [25]

    Dragdiffusion: Harnessing diffusion mod- els for interactive point-based image editing

    Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion mod- els for interactive point-based image editing. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni...

  18. [26]

    As-rigid-as-possible surface modeling

    Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry pro- cessing, pages 109–116. Citeseer, 2007. 3

    2007

  19. [27]

    Least-squares rigid motion using svd

    Olga Sorkine-Hornung and Michael Rabinovich. Least-squares rigid motion using svd. Computing, 1 (1):1–5, 2017. 4

    2017

  20. [28]

    Resolution-robust large mask inpainting with fourier convolutions

    Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Ki- woong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF win...

  21. [29]

    Structured 3d latents for scalable and versatile 3d generation

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. In Proceedings of the Computer Vision and Pattern Recognition Con- ference, pages 21469–21480, 2025. 2, 3

    2025

  22. [30]

    Pixelhacker: Image inpainting with structural and semantic consistency

    Ziyang Xu, Kangsheng Duan, Xiaolei Shen, Zhifeng Ding, Wenyu Liu, Xiaohu Ruan, Xiaoxin Chen, and Xinggang Wang. Pixelhacker: Image inpainting with structural and semantic consistency. arXiv preprint arXiv:2504.20438, 2025. 4

    2025 arXiv

  23. [31]

    Paint by example: Exemplar-based image edit- ing with diffusion models

    Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image edit- ing with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18381–18391, 2023. 1

    2023

  24. [32]

    Magicbrush: A manually annotated dataset for instruction-guided image editing

    Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neu- ral Information Processing Systems, 36:31428–31449,

  25. [33]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF interna- tional conference on computer vision , pages 3836– 3847, 2023. 2 8

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.