Pith. sign in

REVIEW 4 major objections 6 minor 55 references

This paper claims drag-based image editing can be decomposed into pixel-space bidirectional warping followed by standard inpainting, turning any inpainting model into a drag editor that runs in 0.3 seconds at 512×512.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-05 05:56 UTC pith:RQM7IFTC

load-bearing objection Useful method with a real speed/modularity win, but the headline MD comparison is compromised by re-annotated benchmarks. the 4 major comments →

arxiv 2509.04582 v1 pith:RQM7IFTC submitted 2025-09-04 cs.CV

Inpaint4Drag: Repurposing Inpainting Models for Drag-Based Image Editing via Bidirectional Warping

Jingyi Lu , Kai Han This is my paper
classification cs.CV
keywords drag-based image editingimage inpaintingbidirectional warpingpixel-space deformationinteractive editingdiffusion modelselastic deformationimage manipulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Drag-based image editing usually works by nudging the latent space of a generative model, which makes precise, large deformations slow and hard to control. This paper claims the drag interaction can be decomposed into two pixel-space steps: warp the region the user selects, then let a standard inpainting model fill in whatever the deformation revealed. Because the output of the warping is exactly the input format an inpainting model expects, the method works as a drop-in adapter for any existing or future inpainting model. The payoff, if the claim holds, is drag editing at interactive speed—a 0.01-second warp preview and 0.3-second inpainting at 512×512—with precision that matches or beats latent-space methods on the DragBench benchmarks.

Core claim

The paper's central claim is that precise and intuitive drag editing does not require manipulating generative model latents at all. Instead, treating the masked region as a deformable elastic sheet, Inpaint4Drag performs a bidirectional warp: forward warping moves pixels and contours toward the user's targets under inverse-distance weighting, and backward mapping fills every target pixel from nearby forward-matched neighbors so stretching creates no gaps. The result is a warped image plus an inpainting mask covering revealed areas and boundary bands, which is precisely the input of standard inpainting models. The paper reports that this pipeline, with Stable Diffusion inpainting as the backe

What carries the argument

The central mechanism is the bidirectional warping algorithm. Forward warping moves every pixel of the user-selected region by an inverse-distance-weighted sum of the handle-to-target offsets, transforming contours into their target shapes; backward mapping then fills every pixel of the target contour by interpolating from the nearest forward-matched pixels, closing the gaps that pure forward warping leaves in stretched regions. The algorithm outputs a fully covered warped image and a dilation-expanded inpainting mask covering both revealed areas and boundary bands, which together are exactly the input format of a standard inpainting model. The physics framing treats the masked region as an

Load-bearing premise

The reported precision advantage rests on the assumption that the authors' re-annotated masks and drag points on DragBench-S and DragBench-D preserve the original user editing intentions closely enough that MD and LPIPS can be compared head-to-head with numbers previously published on the original annotations.

What would settle it

Run Inpaint4Drag on the original DragBench-S/D annotations exactly as released, without re-annotating masks or points, and compare MD/LPIPS against the published numbers for FastDrag and DragDiffusion; if MD rises above 4.1/5.1 or qualitative dragging on the original point pairs fails, the claimed precision advantage depends on the re-annotation rather than the method.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Users can preview the geometric result of a drag in about 10 milliseconds and iterate masks and handle points before committing to inpainting, changing drag editing from trial-and-error generation to a direct manipulation loop.
  • Because drag inputs are converted to a warped image and mask, any inpainting model can be swapped in without retraining; improvements in inpainting quality should transfer automatically to drag editing.
  • The method handles large occlusions (e.g., opening a lion's mouth, rotating a head) that latent-space drag methods typically fail on, since inpainting models are trained to synthesize missing regions.
  • At 2.7 GB GPU memory and 0.3 seconds at 512×512, drag editing becomes feasible on consumer GPUs and in interactive applications.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A fair head-to-head extension would rerun FastDrag, DragDiffusion, and DiffEditor on the same re-annotated DragBench inputs to remove annotation variation from the comparison; the paper does not do this, so its numerical advantage over prior published scores is partially an inference.
  • The 0.01s preview suggests a slider-based 'scrub' interface where users watch deformation continuously and release at the desired shape; the paper demonstrates multi-round editing but does not evaluate such an interaction.
  • The mask dependency points to an automated region-proposal extension (segmenting object parts by decomposed SAM masks) that could reduce the user's burden and make the method more robust; the paper leaves this as future work.
  • Since the warping is purely geometric and model-agnostic, the same warped input could be fed to a video-inpainting model to propagate edits across frames, an extension the paper does not explore.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes Inpaint4Drag, a drag-based image editing method that operates in pixel space rather than in the latent space of generative models. The user specifies a mask and handle/target point pairs; the method first performs bidirectional warping—forward warping to define target contours and backward mapping to fill gaps—and then passes the warped image and a computed inpainting mask to a standard inpainting model (SD-1.5-Inpaint with LCM/TAESD). The claimed contributions are a real-time warping preview (0.01s), fast inpainting (0.3s at 512x512), a universal adapter for any inpainting model, and an optional SAM-based mask refinement module. Quantitative results on DragBench-S/D and qualitative comparisons against DragDiffusion, DiffEditor, SDE-Drag, and FastDrag are reported.

Significance. If the results hold, the core idea is practically valuable: separating geometric deformation from content generation makes drag editing faster and model-agnostic, and the paper includes a detailed algorithm description, pseudo-code, and a supplementary study with several inpainting backends. The speed advantage is plausible from the decomposition. However, the headline quantitative claims are currently not supported because the evaluation protocol re-annotates the benchmarks in a way that favors the proposed input format, the baseline numbers are not shown to come from controlled re-runs, and a closely related region-based baseline is omitted. The idea is sound enough to warrant a revision, but the evidence needs to be put on an apples-to-apples footing.

major comments (4)
  1. [Sec. 4.1, Table 1] The central MD/LPIPS comparison is not apples-to-apples. The paper states that DragBench-S/D masks and dragging points were 're-annotated' for deformable regions, while the published numbers for DragDiffusion, DiffEditor, SDE-Drag, and FastDrag were produced with the original annotation protocol. The mask is not a neutral input in this method: it directly drives the warping equations (Eqs. 6-12) and the inpainting mask (Eq. 12), so re-annotation can systematically make the task easier for Inpaint4Drag. The claim in Sec. 4.4 of 'lowest MD scores (3.6/3.9)' is therefore not substantiated by Table 1. The authors should re-run all baselines on the same re-annotated masks/points, or provide a per-sample paired comparison with error bars and, ideally, a user study.
  2. [Sec. 4.3/4.4, Table 1] The speed and memory comparison is not shown to be controlled. The manuscript reports Times and GPU memory for baselines, but it does not state that all methods were re-run on the same hardware, software versions, and input resolutions in this work. If these numbers are taken from prior publications, the '600x faster than DragDiffusion' and '14x faster than FastDrag' claims are not reliable because the baselines may have been measured on different GPUs. Please provide a controlled timing benchmark on the same V100 with the same inputs, or clearly state the source of each baseline timing and discuss hardware differences.
  3. [Sec. 4.6, Fig. 7] The key ablation for the central contribution—bidirectional versus unidirectional warping—is only qualitative. Figure 7 shows a single example, and no MD/LPIPS or timing comparison is provided for the two variants on the benchmarks. Since the bidirectional warping mechanism is the main technical novelty, a quantitative ablation on DragBench-S/D (or at least on a representative subset) is necessary to support the claim that it 'fixes sampling gaps' and improves dragging precision.
  4. [Sec. 4.3, Table 1] The comparison omits RegionDrag [22], a region-based diffusion editing method from which the LPIPS/MD evaluation protocol is borrowed. RegionDrag is the closest competitor in terms of the interaction model (region masks plus drag points), so excluding it from Table 1 makes the 'state-of-the-art' claim incomplete. The authors should include RegionDrag in the quantitative comparison or explicitly justify its exclusion.
minor comments (6)
  1. [Sec. 4.2] The paper acknowledges that LPIPS is not an ideal metric for drag editing because correct deformations increase LPIPS, yet it is still used as the primary image-consistency metric in Table 1. Consider adding a user study or additional metrics (e.g., region-specific FID/CLIP scores) and clarifying that LPIPS is a secondary indicator.
  2. [Sec. 4.3] The hyperparameters r1=10, r2=5, Nn=4, epsilon, and |Ps|<=128 are set without a sensitivity analysis. A brief study of their effect on MD/LPIPS and runtime would strengthen the paper.
  3. [Sec. 3.2.1, Eq. (5)] The control-point association rule 'hi inside C' is ambiguous for points exactly on a contour or for nested contours. Please specify how such boundary cases are handled.
  4. [Sec. 3.1, Eq. (3)] The symbol M is reused for the input mask and the refined mask. Rename the initial mask (e.g., M0) to avoid confusion in Eqs. (1)-(3) and Algorithm 1.
  5. [Fig. 5 and Figs. S4-S6] The qualitative comparisons show sparse input for baselines but mask-plus-points for the proposed method. This conflates the interface difference with editing quality. A caption note pointing this out would improve fairness.
  6. [Supplementary Table S1] DeepFillv2 achieves lower MD (3.2/3.7) than the SD-1.5-Inpaint backbone used in the main paper. This is worth a sentence in the main text because it complicates the claim that a specific inpainting model is the best choice.

Circularity Check

0 steps flagged

No significant circularity: the warp/inpaint decomposition is self-contained; re-annotation is a benchmark-validity concern, not a derivation-level circularity.

full rationale

The paper's claimed derivation chain is self-contained. The bidirectional warping algorithm (Eqs. 6–12) is a fixed interpolation scheme driven by explicit user inputs (mask, handle/target points) and standard hyperparameters (ε = 10^-6, Nn = 4, r1 = 10, r2 = 5); no parameter is fitted to DragBench outcomes, and no core result is imported from a self-citation. The inpainting stage (Eq. 13) uses a pretrained external SD-1.5-inpaint checkpoint plus standard accelerators (TAESD, LCM LoRA), so the final edited image is not an encoding of the evaluation metric. The only self-citation is to RegionDrag [22] in Sec. 4.2 for the LPIPS/MD evaluation protocol, which is a standard, non-load-bearing choice. The re-annotation of DragBench-S/D masks and dragging points (Sec. 4.1) is a legitimate benchmark-comparison confound that may affect the fairness of the reported MD/LPIPS numbers versus prior published results, but it is not circularity in the derivation: the re-annotated points are user-style inputs rather than fitted parameters, and the evaluation still relies on an external DIFT matcher. No equation reduces to its own inputs by construction, and no uniqueness claim or ansatz is smuggled in via self-citation. The paper's own limitation statement (S7) concerns sensitivity to user input quality, which is a usability issue, not a circularity issue.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claim rests on a small set of hand-set hyperparameters and four domain assumptions about the deformation model, the interpolation, the evaluation metric, and the benchmark re-annotation. No new physical entities or fitted constants are introduced.

free parameters (6)
  • dilation/erosion radius r1 = 10 px
    Kernel radius for mask refinement in Eq. (3); chosen to provide sufficient boundary refinement while preserving user intent (Sec. 4.3).
  • inpainting mask dilation radius r2 = 5 px
    Dilation kernel radius for the inpainting mask in Eq. (12); set to create a smooth transition buffer (Sec. 4.3).
  • nearest neighbors Nn = 4
    Number of nearest forward-matched pixels used in backward mapping Eq. (8); chosen for computational efficiency and local coherence (Sec. 4.3).
  • diffusion sampling steps = 8
    Number of denoising steps with LCM LoRA for the SD-1.5 inpainting back-end (Sec. 4.3).
  • epsilon for inverse distance weights = 1e-6
    Numerical stabilizer in Eqs. (7) and (8) (Sec. 4.3).
  • SAM sample point limit |Ps| = <=128
    Cap on grid points sampled from the user mask for SAM mask refinement (Sec. 4.3).
axioms (4)
  • domain assumption Image regions act as elastic materials, so inverse-distance-weighted sums of control point displacements give a natural deformation field.
    Introduced in Sec. 3.2.2, Eqs. (6)-(7); this is a modeling choice, not derived from physics.
  • domain assumption Backward mapping using Nn nearest forward-matched pixels provides adequate interpolation to fill holes left by forward warping.
    Assumed in Eq. (8), Sec. 3.2.3; no proof that this inverse interpolation is accurate for large or folding deformations.
  • domain assumption DIFT-based mean distance computed near handle and target points is a valid proxy for drag editing accuracy.
    Used in Sec. 4.2; relies on DIFT correspondences being reliable for the edited images.
  • domain assumption The edited benchmark annotations preserve user intentions that are comparable to the original DragBench annotations.
    Sec. 4.1 re-annotates deformable regions and points to fit the method; comparability with prior published numbers is assumed.

pith-pipeline@v1.4.0-alltime-deepseek-medium · 13256 in / 13008 out tokens · 113759 ms · 2026-08-05T05:56:05.506622+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Inpaint4Drag: Repurposing Inpainting Models for Drag-Based Image Editing via Bidirectional Warping." pith.science (2026). https://pith.science/paper/RQM7IFTC

@misc{pith2026250904582,
  author       = {Pith},
  title        = {Pith review of: Inpaint4Drag: Repurposing Inpainting Models for Drag-Based Image Editing via Bidirectional Warping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RQM7IFTC}},
  note         = {Machine review of arXiv:2509.04582}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Drag-based image editing has emerged as a powerful paradigm for intuitive image manipulation. However, existing approaches predominantly rely on manipulating the latent space of generative models, leading to limited precision, delayed feedback, and model-specific constraints. Accordingly, we present Inpaint4Drag, a novel framework that decomposes drag-based editing into pixel-space bidirectional warping and image inpainting. Inspired by elastic object deformation in the physical world, we treat image regions as deformable materials that maintain natural shape under user manipulation. Our method achieves real-time warping previews (0.01s) and efficient inpainting (0.3s) at 512x512 resolution, significantly improving the interaction experience compared to existing methods that require minutes per edit. By transforming drag inputs directly into standard inpainting formats, our approach serves as a universal adapter for any inpainting model without architecture modification, automatically inheriting all future improvements in inpainting technology. Extensive experiments demonstrate that our method achieves superior visual quality and precise control while maintaining real-time performance. Project page: https://visual-ai.github.io/inpaint4drag/

Figures

Figures reproduced from arXiv: 2509.04582 by Jingyi Lu, Kai Han.

Figure 1
Figure 1. Figure 1: Comparison of drag-based image editing methods. Our approach (rightmost columns) enables both precise edits and large-scale [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our optional mask refinement module. Users can achieve precise object boundaries for coherent defor￾mation through SAM [17], constrained by eroded and dilated mask boundaries to preserve user intent. Drag [28] introduced sparse latent copy-paste techniques, and RegionDrag [22] improved efficiency through dense re￾gion mapping. InstantDrag [42], a motion prediction ap￾proach, combines motion gui… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our bidirectional warping pipeline. Given a user mask and control points, we first extract region contours and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of inpainting mask computation. The final mask ( [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison with state-of-the-art methods on challenging cases from DragBench-S [ [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results of our mask refinement module [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 45 canonical work pages · 1 internal anchor

  1. [1]

    Adobe Photoshop Content-Aware Fill

    Adobe. Adobe Photoshop Content-Aware Fill. https: //www.adobe.com/products/photoshop.html ,

    work page

  2. [2]

    Blended latent diffusion

    Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. TOG, 2023. 2

    work page 2023

  3. [3]

    Patchmatch: A randomized correspon- dence algorithm for structural image editing

    Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspon- dence algorithm for structural image editing. TOG, 2009. 3

    work page 2009

  4. [4]

    Image inpainting

    Marcelo Bertalmio, Guillermo Sapiro, Vincent Caselles, and Coloma Ballester. Image inpainting. In ACM SIGGRAPH,

    work page

  5. [5]

    In- structpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. In- structpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2

    work page 2023

  6. [6]

    Masactrl: Tuning-free mu- tual self-attention control for consistent image synthesis and editing

    Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xi- aohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mu- tual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465, 2023. 2

    Pith/arXiv arXiv 2023

  7. [7]

    Inverting the generator of a generative adversarial network

    Antonia Creswell and Anil Anthony Bharath. Inverting the generator of a generative adversarial network. IEEE TNNLS,

    work page

  8. [8]

    Re- gion filling and object removal by exemplar-based image in- painting

    Antonio Criminisi, Patrick P’erez, and Kentaro Toyama. Re- gion filling and object removal by exemplar-based image in- painting. IEEE TIP, 2004. 3

    work page 2004

  9. [9]

    User-controllable latent transformer for stylegan image layout editing

    Yuki Endo. User-controllable latent transformer for stylegan image layout editing. In CGF, 2022. 2

    work page 2022

  10. [10]

    Diffusion self-guidance for control- lable image generation

    Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for control- lable image generation. In NeurIPS, 2024. 2

    work page 2024

  11. [11]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS,

    work page

  12. [12]

    Easydrag: Efficient point-based manipu- lation on diffusion models

    Xingzhong Hou, Boxiao Liu, Yi Zhang, Jihao Liu, Yu Liu, and Haihang You. Easydrag: Efficient point-based manipu- lation on diffusion models. In CVPR, 2024. 1, 2

    work page 2024

  13. [13]

    Scaling up gans for text-to-image synthesis

    Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In CVPR, 2023. 2

    work page 2023

  14. [14]

    Analyzing and improving the image quality of stylegan

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020

    work page 2020

  15. [15]

    Alias-free generative adversarial networks

    Tero Karras, Miika Aittala, Samuli Laine, Erik H ¨ark¨onen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 2

    work page 2021

  16. [16]

    Imagic: Text-based real image editing with diffusion models

    Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, 2023. 2

    work page 2023

  17. [17]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. In ICCV, 2023. 3

    2023

  18. [18]

    Gan compression: Efficient architectures for interactive conditional gans

    Muyang Li, Ji Lin, Yaoyao Ding, Zhijian Liu, Jun-Yan Zhu, and Song Han. Gan compression: Efficient architectures for interactive conditional gans. In CVPR, 2020. 2

    work page 2020

  19. [19]

    Mat: Mask-aware transformer for large hole image in- painting

    Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image in- painting. In CVPR, 2022. 3

    work page 2022

  20. [20]

    Freedrag: Point tracking is not you need for interactive point-based image editing

    Pengyang Ling, Lin Chen, Pan Zhang, Huaian Chen, and Yi Jin. Freedrag: Point tracking is not you need for interactive point-based image editing. arXiv preprint arXiv:2307.04684, 2023. 1, 2

    Pith/arXiv arXiv 2023

  21. [21]

    Image inpainting for ir- regular holes using partial convolutions

    Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for ir- regular holes using partial convolutions. In ECCV, 2018. 3

    work page 2018

  22. [22]

    Regiondrag: Fast region-based image editing with diffusion models

    Jingyi Lu, Xinghui Li, and Kai Han. Regiondrag: Fast region-based image editing with diffusion models. InECCV,

    work page

  23. [23]

    Readout guidance: Learning con- trol from diffusion features

    Grace Luo, Trevor Darrell, Oliver Wang, Dan B Goldman, and Aleksander Holynski. Readout guidance: Learning con- trol from diffusion features. In CVPR, 2024. 1

    work page 2024

  24. [24]

    Rotationdrag: Point-based image editing with rotated diffusion features

    Minxing Luo, Wentao Cheng, and Jian Yang. Rotationdrag: Point-based image editing with rotated diffusion features. arXiv preprint arXiv:2401.06442, 2024. 2

    Pith/arXiv arXiv 2024

  25. [25]

    Lcm-lora: A universal stable-diffusion acceleration module

    Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolin´ario Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 6

    Pith/arXiv arXiv 2023

  26. [26]

    Dragondiffusion: Enabling drag-style manipula- tion on diffusion models

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

    Pith/arXiv arXiv

  27. [27]

    DiffEditor: Boosting Accuracy and Flexibility on Diffusion-based Image Editing

    Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Diffeditor: Boosting accuracy and flex- ibility on diffusion-based image editing. arXiv preprint arXiv:2402.02583, 2024. 2, 7

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  28. [28]

    The blessing of random- ness: Sde beats ode in general diffusion-based image editing

    Shen Nie, Hanzhong Allan Guo, Cheng Lu, Yuhao Zhou, Chenyu Zheng, and Chongxuan Li. The blessing of random- ness: Sde beats ode in general diffusion-based image editing. arXiv preprint arXiv:2311.01410, 2023. 1, 3, 6, 7

    Pith/arXiv arXiv 2023

  29. [29]

    Taesd: Tiny autoencoder for stable diffusion

    Ollin. Taesd: Tiny autoencoder for stable diffusion. https: //github.com/madebyollin/taesd, 2023. 6

    work page 2023

  30. [30]

    DALL-E 3

    OpenAI. DALL-E 3. https://openai.com/dall- e-3, 2024. Accessed: 2024-02-14. 3

    work page 2024

  31. [31]

    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 SIGGRAPH, 2023. 1, 2, 3, 14

    work page 2023

  32. [32]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6

    work page 2019

  33. [33]

    Styleclip: Text-driven manipulation of stylegan imagery

    Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 2 9

    work page 2021

  34. [34]

    Context encoders: Feature learning by inpainting

    Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016. 3

    work page 2016

  35. [35]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 6, 12

    work page 2022

  36. [36]

    RunwayML

    Runway. RunwayML. https://runwayml.com, 2024. Accessed: 2024-02-14. 3

    work page 2024

  37. [37]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2

    work page 2022

  38. [38]

    Image de- formation using moving least squares

    Scott Schaefer, Travis McPhail, and Joe Warren. Image de- formation using moving least squares. In ACM SIGGRAPH,

    work page

  39. [39]

    Free-form deforma- tion of solid geometric models

    Thomas W Sederberg and Scott R Parry. Free-form deforma- tion of solid geometric models. In ACM SIGGRAPH, 1986. 1

    work page 1986

  40. [40]

    Dragdiffusion: Harnessing diffu- sion models for interactive point-based image editing

    Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vin- cent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffu- sion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 1, 2, 6, 7, 8

    Pith/arXiv arXiv 2023

  41. [41]

    Instadrag: Lightning fast and accurate drag- based image editing emerging from videos

    Yujun Shi, Jun Hao Liew, Hanshu Yan, Vincent YF Tan, and Jiashi Feng. Instadrag: Lightning fast and accurate drag- based image editing emerging from videos. arXiv preprint arXiv:2405.13722, 2024. 3, 14

    Pith/arXiv arXiv 2024

  42. [42]

    Instant- drag: Improving interactivity in drag-based image editing

    Joonghyuk Shin, Daehyeon Choi, and Jaesik Park. Instant- drag: Improving interactivity in drag-based image editing. TOG, 2024. 3

    work page 2024

  43. [43]

    Resolution-robust large mask inpainting with fourier convolutions

    Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, 2022. 2, 3, 12

    work page 2022

  44. [44]

    Emergent correspondence from image diffusion

    Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. In NeurIPS, 2024. 6

    work page 2024

  45. [45]

    High-fidelity gan inversion for image attribute editing

    Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In CVPR, 2022. 2

    work page 2022

  46. [46]

    Gan inversion: A survey

    Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. IEEE TPAMI, 2022

    work page 2022

  47. [47]

    Gan prior embedded network for blind face restoration in the wild

    Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In CVPR, 2021. 2

    work page 2021

  48. [48]

    Generative image inpainting with contex- tual attention

    Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contex- tual attention. In CVPR, 2018. 3

    work page 2018

  49. [49]

    Free-form image inpainting with gated convolution

    Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In ICCV, 2019. 12

    work page 2019

  50. [50]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6

    2018

  51. [51]

    Efficientvit-sam: Accelerated segment anything model without performance loss

    Zhuoyang Zhang, Han Cai, and Song Han. Efficientvit-sam: Accelerated segment anything model without performance loss. In CVPR, 2024. 6

    work page 2024

  52. [52]

    Fastdrag: Manipulate anything in one step

    Xuanjia Zhao, Jian Guan, Congyi Fan, Dongli Xu, Youtian Lin, Haiwei Pan, and Pengming Feng. Fastdrag: Manipulate anything in one step. In NeurIPS, 2024. 1, 3, 7, 14

    work page 2024

  53. [53]

    Generative visual manipulation on the natu- ral image manifold

    Jun-Yan Zhu, Philipp Kr ¨ahenb¨uhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natu- ral image manifold. In ECCV, 2016. 2 10 Inpaint4Drag: Repurposing Inpainting Models for Drag-Based Image Editing via Bidirectional Warping – Supplementary Material – Jingyi Lu Kai Han † Visual AI Lab, The University of Hong Kong lujingyi@conn...

    work page 2016

  54. [55]

    While quantitative metrics in Tab

    to recent generative model-based techniques [35]. While quantitative metrics in Tab. S1 show comparable drag editing performance across methods, qualitative differences emerge in Fig. S1. Early approaches offer computational efficiency, whereas generative methods sometimes produce more realistic results–a quality distinction not fully captured by existing...

    work page

  55. [2024]

    Accessed: 2024-02-14. 3

    work page 2024