Pith. sign in

REVIEW 4 major objections 5 minor 57 references

Mastering Regional 3DGS: Locating, Initializing, and Editing with Diverse 2D Priors

T0 review · 4 major / 5 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 A 3D Gaussian Splatting editing pipeline that combines 2D diffusion localization, depth-based point seeding, and sequential view refinement to achieve up to 4x faster local edits.

desk verdict A competent 3DGS editing systems paper whose integration of 2D localization, depth-based seeding, and sequential refinement is new, but whose SOTA and speedup headline is under-supported by a three-scene evaluation and an untested depth-ratio heuristic. read the letter →

arxiv 2507.05426 v1 pith:ZZDG3SWI submitted 2025-07-07 cs.CV

Lanqing Guo , Yufei Wang , Hezhen Hu , Yan Zheng , Yeying Jin , Siyu Huang , Zhangyang Wang This is my paper
classification cs.CV
keywords editingsceneeditsinitializinglocalregionalregionsview
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's goal is to change just one part of a 3D scene, for example adding a mustache to a person or a rose into a vase, while leaving the rest untouched. The scene is stored as a cloud of small Gaussian blobs, and the challenge is to find which blobs to change and then to make the change look right from every camera angle.

The method has three stages. First, it renders several 2D views and feeds them to InstructPix2Pix, an image diffusion model trained to follow edit instructions. By comparing the model's internal noise estimates with and without the instruction, the system builds a 2D map of the pixels that should change. These 2D masks are projected back into the 3D cloud, giving a 3D editing mask.

Second, some edits create geometry that was not there before, like a hat or a mustache. The system takes the edited front view and asks Depth Anything, a monocular depth estimator, to predict how far away each pixel is. It calibrates this depth prediction against the depth that the original scene already renders, then places new Gaussian blobs at the locations of the edited pixels. This rough seeding makes later fine-tuning much faster.

Third, the system refines neighboring views one by one. Each nearby view is re-edited by InstructPix2Pix while conditioning on both the already edited front view and the original unedited view, then the whole 3D scene is fine-tuned to match all edited views. The process repeats with progressively less aggressive diffusion steps.

The authors report about two minutes of editing on one GPU, up to 4 times faster than prior 3D Gaussian editing methods, with competitive or better fidelity on their chosen scenes.

Extended reading notes

Core claim

The central claim, from the abstract, is that the method 'achieves state-of-the-art performance while delivering up to a 4x speedup' for 3D scene local editing. If correct, the pipeline yields edit fidelity comparable or superior to existing 3DGS editing methods (GaussianEditor, DGE, GaussCtrl) while running diffusion-based refinement and 3DGS fine-tuning in roughly two minutes on one A6000 GPU.

Load-bearing premise

The depth-based initialization (Sec. 4.2) assumes that the ratio of Depth Anything predictions on the edited and unedited frontal views, d_mono(x_edited)/d_mono(x_unedit), multiplied by the existing rendered depth d_3dgs, places newly added Gaussians (mustaches, hats, glasses) at their true 3D location, so they render correctly from other angles (Eq. 7). This is a heuristic: the depth model never saw the edited image during training, and a pure color/texture edit should produce no geometric depth change, so any false geometry signal in this ratio will misplace points and cause the exact side-view inconsistencies that the ablation shows when initialization is removed.

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 / 5 minor

Summary. The paper proposes a pipeline for localized editing of 3D Gaussian Splatting scenes. It localizes the edit region by comparing IP2P noise predictions with and without the text prompt across rendered views, inverse-renders the resulting 2D masks to a 3D mask, initializes newly added geometry using Depth Anything depth maps from the edited frontal view, and then refines adjacent views through conditional IP2P editing interleaved with 3DGS finetuning. Experiments on three scenes report CTIDS, IIS, FID, and editing time against NeRF- and 3DGS-based baselines, complemented by qualitative ablations of localization and initialization and a user study. The central claim is state-of-the-art local editing with up to 4x speedup.

Significance. If the results hold, the main contribution is a practical coarse-to-fine framework that separates localization, initialization, and refinement, and the qualitative results are often visually cleaner than the baselines. The idea of using a depth ratio to seed added Gaussians is novel and potentially useful. The paper also demonstrates a large speedup over prior editing pipelines and a user-study preference. However, the evidence base is currently too narrow: only three scenes are evaluated, there is no per-scene quantitative breakdown, no error bars, no quantitative ablation of the two claimed contributions, and the depth-ratio initialization rests on an untested assumption about Depth Anything's behavior on edited images. The paper has the ingredients of a solid systems contribution, but the evaluation must be substantially strengthened before the state-of-the-art and speedup claims can be accepted.

major comments (4)
  1. [§5.4, Table 1; §5.6] The quantitative evaluation is too thin to support the stated claims. Table 1 reports aggregate metrics over only three scenes, with no per-scene breakdown, no error bars, no statistical significance, and no quantitative ablation of the localization and initialization components. The abstract's 'state-of-the-art' claim is not supported even by Table 1 itself: Ours achieves CTIDS 0.2053, which is below DGE's 0.2057, while IIS and FID are better; with a single aggregate number, this is not enough to establish superiority. Please provide a full per-scene table with standard deviations, a quantitative ablation (e.g., removing localization or initialization and reporting CTIDS/IIS/FID), and a description of the user-study protocol.
  2. [§4.2, Eqs. (6)-(7); §5.6, Fig. 8; Limitations] The depth-ratio initialization assumes that Depth Anything's prediction on the IP2P-edited frontal view is geometrically meaningful. The scale a and bias b in Eq. (6) are calibrated on the unedited 3DGS-rendered depth and then applied to the edited image. For a pure color/texture edit, the true depth is unchanged, so any edit-induced change in d_mono(x_edited)/d_mono(x_unedit) is spurious and can displace the added Gaussians. The ablation in Fig. 8 shows only that removing initialization hurts side-view consistency, not that the depth-ratio placement is correct; the Limitations section lists lighting, low texture, reflections, and occlusions as failure modes but does not address edit-induced out-of-distribution inputs. Please validate the initialization, for example on scenes with known geometry or against a manual/ground-truth 3D mask, and report the resulting 3D placement error.
  3. [§4.1, Eq. (5); §4.3] The 2D localization mask and the edited views are both produced by the same IP2P model. This is a pragmatic co-design rather than a logical circularity, but it means the localization is evaluated only in combination with the editor that generated the mask, so the reported quality does not establish that the 3D localization is editor-agnostic. Please test with at least one different 2D editor or with ground-truth/localization masks, and discuss how mask errors from a different prior affect the final result.
  4. [§5.4, Table 1; §5.5] The speedup claim needs a precise timing protocol. The manuscript does not state whether the reported editing times were measured on the same GPU and implementation for all baselines, whether they include all stages (localization, initialization, refinement, finetuning), or how the numbers were averaged over the three scenes. The user study is also reported only as a single voting percentage, with no details on the number of scenes, the questions asked, or statistical significance. Please provide the timing breakdown and the user-study materials.
minor comments (5)
  1. [§4.2, Eq. (7)] The notation in Eq. (7) is unclear: M is introduced as a 3D mask, but the right-hand side P maps 2D pixels to 3D points; please clarify the domain of M and the role of x_edit in P.
  2. [§4.1, Eq. (5)] The function F (threshold plus Gaussian filtering) is not specified precisely; please give the exact thresholding and filter parameters.
  3. [§4.2] The sentence 'we propose to randomly select a few views' at the beginning of Section 4.2 conflicts with the deterministic choice vfirst = argmax_v sum_{i,j} M_{v,i,j}; please rephrase to avoid the apparent contradiction.
  4. [§5.1] The statement that edit strength is 'adjusted consistently across all competing methods' for certain cases is too vague to reproduce; please list the per-case settings.
  5. [§5.4] Please specify which rendered views are used for the CTIDS/IIS/FID computations and whether they are identical across all compared methods.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the 3D localization, depth-ratio initialization, and sequential refinement are linked by shared 2D priors but not by definitional or fitted-input reductions.

full rationale

The derivation chain is self-contained rather than circular. Equation 5 defines the 2D localization mask M_v as a text-vs-empty IP2P noise-difference map, while the edited frontal view is also generated by IP2P; sharing the same 2D editor as both localizer and editor is a pragmatic coupling, not a definitional reduction of one quantity to the other. Equation 6 fits the depth-calibration scale a and bias b to the unedited 3DGS rendered depth, and Equation 7 then uses the calibrated monocular depth ratio d_mono(x_edited)/d_mono(x_unedit) times d_3dgs to place new Gaussians; this is a scene-level calibration reused for initialization, not a parameter fitted to the target edit or to the evaluation metric. The final output is produced by the LPIPS/L1 3DGS optimization in Equation 9, so the reported CTIDS/IIS/FID numbers in Table 1 are measured on rendered edited views rather than being forced by the equations. The only self-citation in a potentially load-bearing position is [55] for the general claim that good initialization is important in sparse-view settings, but the paper independently ablates initialization in Figures 8 and 9, so that citation is not load-bearing. The Limitations paragraph honestly lists depth-failure regimes such as extreme lighting, low texture, reflections, and occlusions; it does not separately discuss edit-induced out-of-distribution depth, but that is a correctness risk rather than a circularity. No step reduces by construction to its own input or to a fitted value renamed as a prediction.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central contribution is a system-level pipeline; its claims rest on the reliability of pretrained 2D models (IP2P, Depth Anything), a scene-specific depth calibration, and seven hand-set hyperparameters. No new physical or semantic entities are introduced beyond the standard Gaussian representation.

free parameters (7)
  • Localization timestep tau = 600
    Timestep in Eq. 5 where the noise-difference mask is computed; chosen by hand, no sensitivity analysis.
  • Mask threshold gamma = 0.6
    Threshold in F() to binarize the noise-difference map; chosen by hand.
  • Classifier-free guidance strength = 7.5
    Guidance weight w in Eq. 3 for IP2P; default value, adjusted for some scenes without a sensitivity study.
  • Starting denoising steps t per cycle = [750,500,250] (3 cycles) or [750,500] (2 cycles)
    Linearly decreasing starting timesteps for IP2P refinement; chosen by hand to trade semantic alignment for detail.
  • Number of adjacent views m = 20
    Number of neighboring views refined per cycle in Eq. 8; chosen by hand.
  • Depth calibration scale a and bias b = per-scene values from Eq. 6
    Fitted to align Depth Anything disparity to the 3DGS rendered depth on unedited views; used to unproject edited pixels in Eq. 7.
  • 3DGS finetuning iterations = 1000 (kitchen) to 1500 (others)
    Number of Gaussian optimization steps per scene, chosen to balance speed and convergence.
assumptions (4)
  • domain assumption IP2P noise-difference masking identifies all pixels relevant to the edit, including content that must be added.
    The difference between predicted noises with and without the prompt (Eq. 5) must yield a complete and accurate region. If it misses part of the target or fires on distractors, the 3D mask is wrong; the paper provides no formal justification.
  • ad hoc to paper Depth Anything predictions on the IP2P-edited frontal view are geometrically meaningful.
    The ratio d_mono(edited)/d_mono(unedited) in Eq. 7 treats monocular depth on an edited, out-of-distribution image as geometric truth. This is a heuristic central to the initialization stage.
  • domain assumption A single affine map (a,b) from monocular disparity to 3DGS depth holds across the whole scene.
    Eq. 6 assumes one global scale and bias align Depth Anything disparity to the scene's metric depth. This can break at object boundaries or on reflective surfaces, which the paper itself lists as limitations.
  • domain assumption Optimizing all Gaussians with L1 and LPIPS loss against the edited views yields a consistent 3D scene.
    Eq. 9 relies on the edited views being mutually compatible. The sequential refinement is designed to create compatibility, but the loss assumes it already holds sufficiently for optimization to converge.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mastering Regional 3DGS: Locating, Initializing, and Editing with Diverse 2D Priors." pith.science (2026). https://pith.science/paper/ZZDG3SWI

@misc{pith2026250705426,
  author       = {Pith},
  title        = {Pith review of: Mastering Regional 3DGS: Locating, Initializing, and Editing with Diverse 2D Priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZZDG3SWI}},
  note         = {Machine review of arXiv:2507.05426}
}
abstract

Many 3D scene editing tasks focus on modifying local regions rather than the entire scene, except for some global applications like style transfer, and in the context of 3D Gaussian Splatting (3DGS), where scenes are represented by a series of Gaussians, this structure allows for precise regional edits, offering enhanced control over specific areas of the scene; however, the challenge lies in the fact that 3D semantic parsing often underperforms compared to its 2D counterpart, making targeted manipulations within 3D spaces more difficult and limiting the fidelity of edits, which we address by leveraging 2D diffusion editing to accurately identify modification regions in each view, followed by inverse rendering for 3D localization, then refining the frontal view and initializing a coarse 3DGS with consistent views and approximate shapes derived from depth maps predicted by a 2D foundation model, thereby supporting an iterative, view-consistent editing process that gradually enhances structural details and textures to ensure coherence across perspectives. Experiments demonstrate that our method achieves state-of-the-art performance while delivering up to a $4\times$ speedup, providing a more efficient and effective approach to 3D scene local editing.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 41 canonical work pages

  1. [1]

    Renderdiffu- sion: Image diffusion for 3d reconstruction, inpainting and generation

    Titas Anciukeviˇ cius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffu- sion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 12608– 12618, 2023

    work page 2023

  2. [2]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hed- man. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 5470–5479, 2022

    work page 2022

  3. [3]

    Tim Brooks, Arya Ghassemi, Aleksander Holynski, and Alexei A. Efros. Grounding language models to images for multi-region editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2023

    work page 2023

  4. [4]

    Instructpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 18392–18402, 2023

    work page 2023

  5. [5]

    Dge: Direct gaussian 3d editing by consistent multi-view editing

    Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Direct gaussian 3d editing by consistent multi-view editing. arXiv preprint arXiv:2404.18929, 2024

    arXiv 2024

  6. [6]

    Upst-nerf: Universal pho- torealistic style transfer of neural radiance fields for 3d scene

    Yaosen Chen, Qi Yuan, Zhiqiang Li, Yue- gen Liu, Wei Wang, Chaoping Xie, Xuming Wen, and Qien Yu. Upst-nerf: Universal pho- torealistic style transfer of neural radiance fields for 3d scene. IEEE Transactions on Visualization and Computer Graphics , 2024

    work page 2024

  7. [7]

    Gaussianeditor: Swift and controllable 3d editing with gaussian splatting

    Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 21476–21485, 2024

    work page 2024

  8. [8]

    Vica- nerf: View-consistency-aware 3d editing of neural radiance fields

    Jiahua Dong and Yu-Xiong Wang. Vica- nerf: View-consistency-aware 3d editing of neural radiance fields. Advances in Neural Information Processing Systems, 36, 2024

    work page 2024

Show all 57 references
  1. [9]

    threestudio: A unified framework for 3d content generation

    Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zi-Xin Zou, Chen Wang, Yan-Pei Cao, et al. threestudio: A unified framework for 3d content generation. three- studio: A unified framework for 3d content generation, 2023

    2023

  2. [10]

    Instruct-nerf2nerf: Editing 3d scenes with instructions

    Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 19740–19750, 2023. 13

    2023

  3. [11]

    Tenen- baum, Antonio Torralba, David Bau, and Tali Dekel

    Amir Hertz, Ron Mokady, Dan B. Tenen- baum, Antonio Torralba, David Bau, and Tali Dekel. Prompt-to-prompt image editing with cross-attention control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 5291–5300, 2022

    2022

  4. [12]

    Gans trained by a two time-scale update rule converge to a local nash equi- librium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equi- librium. In Advances in Neural Information Processing Systems, pages 6626–6637, 2017

    2017

  5. [13]

    Learning locally editable virtual humans

    Hsuan-I Ho, Lixin Xue, Jie Song, and Otmar Hilliges. Learning locally editable virtual humans. In Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  6. [14]

    Classifier- free diffusion guidance

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

    2022 arXiv

  7. [15]

    Image-to-image transla- tion with conditional adversarial networks

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image transla- tion with conditional adversarial networks. In Proceedings of the IEEE conference on com- puter vision and pattern recognition (CVPR), pages 1125–1134, 2017

    2017

  8. [16]

    Spatial transformer networks

    Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In Advances in Neural Information Processing Systems (NeurIPS) , pages 2017–2025, 2015

    2017

  9. [17]

    Stylesplat: 3d object style transfer with gaussian splatting

    Sahil Jain, Avik Kuthiala, Prabhdeep Singh Sethi, and Prakanshul Saxena. Stylesplat: 3d object style transfer with gaussian splatting. arXiv preprint arXiv:2407.09473 , 2024

    2024 arXiv

  10. [18]

    Free- man, and Jonathan Ho

    Bahjat Kawar, Huiwen Chang, Tim Sali- mans, Phillip Isola, Ido Mosseri, Peyman Milanfar, Jeffrey De Fauw, William T. Free- man, and Jonathan Ho. Imagic: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276 , 2023

    arXiv

  11. [19]

    3d gaus- sian splatting for real-time radiance field rendering

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

    2023

  12. [20]

    3d gaus- sian splatting for real-time radiance field ren- dering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk¨ uhler, and George Drettakis. 3d gaus- sian splatting for real-time radiance field ren- dering. ACM Transactions on Graphics , 42 (4), July 2023. URL https://repo-sam.inria. fr/fungraph/3d-gaussian-splatting/

    2023

  13. [21]

    A hierarchical 3d gaussian representation for real-time ren- dering of very large datasets

    Bernhard Kerbl, Andreas Meuleman, Geor- gios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian representation for real-time ren- dering of very large datasets. ACM Transac- tions on Graphics (TOG) , 43(4):1–15, 2024

    2024

  14. [22]

    Repaint: Inpainting using denois- ing diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Fabian Timm, and Luc Van Gool. Repaint: Inpainting using denois- ing diffusion probabilistic models. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11461–11471, 2022

    2022

  15. [23]

    Sdedit: Image synthesis and editing with stochas- tic differential equations

    Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochas- tic differential equations. arXiv preprint arXiv:2108.01073, 2021

    2021 arXiv

  16. [24]

    Local light field fusion: Practical view syn- thesis with prescriptive sampling guidelines

    Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view syn- thesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG) , 38 (4):1–14, 2019

    2019

  17. [25]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Represent- ing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV) , pages 405–421, 2020

    2020

  18. [26]

    Watch your steps: Local image and scene editing by text instructions

    Ashkan Mirzaei, Tristan Aumentado- Armstrong, Marcus A Brubaker, Jonathan Kelly, Alex Levinshtein, Konstantinos G Derpanis, and Igor Gilitschenski. Watch your steps: Local image and scene editing by text instructions. In European Conference on Computer Vision , pages 111–129. ...

    2025

  19. [27]

    Null-text inversion for editing real images using guided diffusion models

    Ron Mokady, Amir Hertz, Or Perel, Kfir Aberman, Dani Lischinski, and Tali Dekel. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2301.09725, 2023

    2023 arXiv

  20. [28]

    Zero-shot text-guided object manipulation 14 in diffusion latent space

    Gaurav Parmar, Patrick Esser, Yunjey Zhang, Boris Dayma, Angelos Katharopou- los, Gabriel Synnaeve, and Yann LeCun. Zero-shot text-guided object manipulation 14 in diffusion latent space. arXiv preprint arXiv:2303.13078, 2023

    2023 arXiv

  21. [29]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis

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

    2023 arXiv

  22. [30]

    Deep learning- based image and video inpainting: A survey

    Weize Quan, Jiaxi Chen, Yanli Liu, Dong- Ming Yan, and Peter Wonka. Deep learning- based image and video inpainting: A survey. International Journal of Computer Vision , 132(7):2367–2400, 2024

    2024

  23. [31]

    Learning trans- ferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hal- lacy, Aditya Ramesh, Gabriel Goh, Sand- hini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning trans- ferable visual models from natural language supervision. In Proceedings...

    2021

  24. [32]

    Dreambooth: Fine tuning text- to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text- to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2250...

    2023

  25. [33]

    Magic insert: Style-aware drag-and-drop

    Nataniel Ruiz, Yuanzhen Li, Neal Wadhwa, Yael Pritch, Michael Rubinstein, David E Jacobs, and Shlomi Fruchter. Magic insert: Style-aware drag-and-drop. arXiv preprint arXiv:2407.02489, 2024

    2024 arXiv

  26. [34]

    A survey of multimodal-guided image editing with text-to-image diffusion models

    Xincheng Shuai, Henghui Ding, Xingjun Ma, Rongcheng Tu, Yu-Gang Jiang, and Dacheng Tao. A survey of multimodal-guided image editing with text-to-image diffusion models. arXiv preprint arXiv:2406.14555 , 2024

    2024 arXiv

  27. [35]

    Splicing vit fea- tures for semantic appearance transfer

    Edgar Tumanyan, Fangneng Zhan, Jian Chen, Varun Jampani, Qiang Liu, Marc Pollefeys, and Angela Dai. Splicing vit fea- tures for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13335–13344, 2022

    2022

  28. [36]

    Pix2pix- zero: Text-guided image-to-image transla- tion without forgetting

    Yaniv Valevski, Edgar Tumanyan, Sergey Tulyakov, and Anat Levin Noy. Pix2pix- zero: Text-guided image-to-image transla- tion without forgetting. arXiv preprint arXiv:2302.03027, 2023

    2023 arXiv

  29. [37]

    Nerf-art: Text-driven neural radiance fields stylization

    Can Wang, Ruixiang Jiang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics , 2023

    2023

  30. [38]

    Exploiting diffusion prior for real-world image super-resolution

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Jour- nal of Computer Vision , 132(12):5929–5949, 2024

    2024

  31. [39]

    Gaussianedi- tor: Editing 3d gaussians delicately with text instructions

    Junjie Wang, Jiemin Fang, Xiaopeng Zhang, Lingxi Xie, and Qi Tian. Gaussianedi- tor: Editing 3d gaussians delicately with text instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20902–20911, 2024

    2024

  32. [40]

    Lavie: High-quality video gener- ation with cascaded latent diffusion models

    Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video gener- ation with cascaded latent diffusion models. International Journal of Computer Vision , 133(5):3059–3078, 2025

    2025

  33. [41]

    Training-free text-guided image edit- ing with visual autoregressive model

    Yufei Wang, Lanqing Guo, Zhihao Li, Jiaxing Huang, Pichao Wang, Bihan Wen, and Jian Wang. Training-free text-guided image edit- ing with visual autoregressive model. arXiv preprint arXiv:2503.23897, 2025

    2025 arXiv

  34. [42]

    View-consistent 3d editing with gaussian splatting

    Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaussian splatting. In European Conference on Com- puter Vision , pages 404–420. Springer, 2025

    2025

  35. [43]

    Ner- filler: Completing scenes via generative 3d inpainting

    Ethan Weber, Aleksander Holynski, Varun Jampani, Saurabh Saxena, Noah Snavely, Abhishek Kar, and Angjoo Kanazawa. Ner- filler: Completing scenes via generative 3d inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20731–20741, 2024

    2024

  36. [44]

    GaussCtrl: Multi-View Con- sistent Text-Driven 3D Gaussian Splatting Editing

    Jing Wu, Jia-Wang Bian, Xinghui Li, Guan- grun Wang, Ian Reid, Philip Torr, and Vic- tor Prisacariu. GaussCtrl: Multi-View Con- sistent Text-Driven 3D Gaussian Splatting Editing. ECCV, 2024

    2024

  37. [45]

    Fast- composer: Tuning-free multi-subject image 15 generation with localized attention

    Guangxuan Xiao, Tianwei Yin, William T Freeman, Fr´ edo Durand, and Song Han. Fast- composer: Tuning-free multi-subject image 15 generation with localized attention. Inter- national Journal of Computer Vision , pages 1–20, 2024

    2024

  38. [46]

    Localized gaussian splatting editing with contextual awareness

    Hanyuan Xiao, Yingshu Chen, Huajian Huang, Haolin Xiong, Jing Yang, Pratusha Prasad, and Yajie Zhao. Localized gaussian splatting editing with contextual awareness. arXiv preprint arXiv:2408.00083 , 2024

    2024 arXiv

  39. [47]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceed- ings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 10371–10381, 2024

    2024

  40. [48]

    Depth anything v2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Heng- shuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024

    2024 arXiv

  41. [49]

    Show-1: Marrying pixel and latent diffusion models for text-to-video generation

    David Junhao Zhang, Jay Zhangjie Wu, Jia- Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision , pages 1–15, 2024

    2024

  42. [50]

    Adding conditional control to text-to-image diffusion models, 2023

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023

    2023

  43. [51]

    The unrea- sonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unrea- sonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018

    2018

  44. [52]

    Oscillation inversion: Under- stand the structure of large flow model through the lens of inversion method

    Yan Zheng, Zhenxiao Liang, Xiaoyan Cong, Yuehao Wang, Peihao Wang, Zhangyang Wang, et al. Oscillation inversion: Under- stand the structure of large flow model through the lens of inversion method. arXiv preprint arXiv:2411.11135, 2024

    2024 arXiv

  45. [53]

    Instructhumans: Editing animated 3d human textures with instructions

    Jiayin Zhu, Linlin Yang, and Angela Yao. Instructhumans: Editing animated 3d human textures with instructions. arXiv preprint arXiv:2404.04037, 2024

    2024

  46. [54]

    Genera- tive visual manipulation on the natural image manifold

    Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Genera- tive visual manipulation on the natural image manifold. European Conference on Computer Vision (ECCV) , pages 597–613, 2016

    2016

  47. [55]

    Fsgs: Real-time few-shot view synthesis using gaussian splatting

    Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision , pages 145–163. Springer, 2025

    2025

  48. [56]

    Dreamedi- tor: Text-driven 3d scene editing with neural fields

    Jingyu Zhuang, Chen Wang, Liang Lin, Lingjie Liu, and Guanbin Li. Dreamedi- tor: Text-driven 3d scene editing with neural fields. In SIGGRAPH Asia 2023 Conference Papers, pages 1–10, 2023

    2023

  49. [57]

    Fmgs: Foun- dation model embedded 3d gaussian splatting for holistic 3d scene understanding

    Xingxing Zuo, Pouya Samangouei, Yunwen Zhou, Yan Di, and Mingyang Li. Fmgs: Foun- dation model embedded 3d gaussian splatting for holistic 3d scene understanding. Interna- tional Journal of Computer Vision , 133(2): 611–627, 2025. 16

    2025

Pith tools

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