Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

2D Gaussian Splatting with Semantic Alignment for Image Inpainting

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that image inpainting can be done by encoding a masked image into a 2D Gaussian splat field and rasterizing the complete image, with adapted DINO features keeping the result semantically coherent.

desk verdict A plausible new application of 2D Gaussian Splatting to inpainting, with good LPIPS and speed, but the semantic alignment evidence is thinner than the claims and the FID overclaim needs scrubbing. read the letter →

arxiv 2509.01964 v1 pith:SCFOTKGR submitted 2025-09-02 cs.CV cs.AI

Hongyu Li , Chaofeng Chen , Xiaoming Li , Guangming Lu This is my paper
classification cs.CVcs.AI
keywords imageinpainting2DGaussiansplattingdifferentiablerasterizationsemanticalignmentDINOv2featuresadaptivelayernormalizationpatch-levelrenderingU-Netencoder
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 tries to establish that a new family of inpainting models can work by treating a masked image not as pixels to be repaired but as the source for a continuous 2D Gaussian splat field, which is then rasterized into the completed image. It claims this is the first use of 2D Gaussian Splatting for inpainting, and that the continuous rendering paradigm gives pixel-level coherence that discrete CNN and Transformer decoders struggle to match. To keep the result globally coherent, the paper adapts DINOv2 features, which it observes stay meaningful under small masks, with a lightweight MLP and injects them through AdaLN into the encoder. Patch-level rasterization with overlapping borders makes the approach memory-efficient and fast while preserving smoothness across patches. If these claims hold, masked images can be completed with both local texture continuity and global semantic consistency in a single feed-forward render, without slow iterative diffusion.

What carries the argument

The central object is the 2D Gaussian splat field: each Gaussian carries a mean position, a color, a coefficient, and a covariance matrix factorized as L L^T to stay positive semi-definite. Pixel values are computed as the sum of weighted Gaussian responses, so the image lives in a continuous, overlapping field rather than on a discrete pixel grid. The method wraps this in a patch-level rasterizer that renders each patch over an extended area and blends overlapping borders, and a semantic alignment module that takes frozen DINOv2 features from the masked image, maps them through a lightweight MLP toward clean feature space, and injects them into every bottleneck layer of a U-Net through AdaL

What would settle it

Take a held-out image with a scene layout unlike CelebA-HQ or Places2 (for example, dense text, repeated identical objects, or a single dominant object), mask 40 to 60 percent of it in an unusual shape, run the trained model, and measure the cosine similarity between the adapted DINO feature and the clean image's DINO feature. If the adapted similarity gain vanishes and the completion inserts semantically wrong content, the central semantic-alignment claim is refuted.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that image inpainting can be reformulated as estimating a complete 2D Gaussian parameter field from a masked image, then rasterizing that field back into pixels; the continuous, overlapping nature of the Gaussians supplies local coherence, while adapted DINO features supply global semantic consistency. The paper supports this with an encoder that outputs Gaussian parameters (mean offsets, colors, covariance factors, coefficients), a patch-wise rasterizer with overlap blending, and an MLP plus AdaLN module that maps masked DINOv2 features toward clean semantic features under an alignment loss. It reports that this configuration reaches FID and LP

Load-bearing premise

The load-bearing premise is that a lightweight MLP can convert DINOv2 features extracted from heavily masked images (40 to 60 percent holes) into clean semantic guidance; if that mapping saturates or fails outside the benchmarked mask shapes and scene categories, the global semantic consistency of the completions has no support.

Editorial extensions

If this is right

  • Inpainting can be reframed as Gaussian parameter estimation plus rendering, so a single feed-forward pass replaces iterative or autoregressive pixel synthesis.
  • Because rendering is continuous and overlapping, completed regions inherit smooth intensity transitions, which shows up as lower LPIPS than CNN-decoder alternatives in the paper's ablations.
  • Patch-wise rasterization with overlap blending lets the approach scale to high resolution and run at about 32.52 ms per image, roughly 2,400 times faster than the RePaint diffusion baseline reported in the paper.
  • Adapted DINO features act as a global semantic prior that can be injected through AdaLN, so semantic coherence is maintained even when processing is patch-local.
  • Removing the DINO adaptation or replacing the Gaussian decoder with a CNN decoder measurably hurts FID, LPIPS, and training stability, supporting the claim that both components carry the result.

Reading between the lines

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

  • The same encoder-rasterization design could be applied to other per-pixel restoration tasks such as dehazing, denoising, or deblurring by changing only the loss and the mask condition; the paper stops at inpainting but frames the 2D Gaussian field as a general continuous image prior.
  • Because the output is an explicit set of Gaussian parameters rather than a latent code, users could edit the inpainted result by moving, deleting, or duplicating individual Gaussians; the paper does not explore this interactivity.
  • The feature-adaptation observation suggests DINOv2 or similar self-supervised features can act as a cheap universal semantic regularizer for masked-image tasks; a direct test would be fine-tuning the same MLP block on object-centric datasets and measuring whether alignment generalizes to unseen categories.
  • Combining the fast 2DGS renderer with a diffusion prior could close the reported FID gap to LaMa while keeping most of the speed; nothing in the paper rules out this hybrid, and the paper explicitly leaves generative-model integration for future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a 2D Gaussian Splatting (2DGS) framework for image inpainting. A masked input is encoded by a U-Net into 2D Gaussian parameters, which are rasterized patch-wise to reconstruct the full image. To improve global semantic consistency, DINOv2 features are extracted from the masked input, adapted by a lightweight MLP, and injected into the bottleneck via Adaptive Layer Normalization (AdaLN), with a cosine-similarity alignment loss to clean features. The method is evaluated on CelebA-HQ and Places2 against several inpainting baselines, reporting competitive LPIPS and generally higher FID than LaMa. Ablations on ImageNet-100 assess the semantic module, the Gaussian decoder, the number of Gaussians, and the AdaLN conditioning.

Significance. If the claims hold, this is a new application direction for 2D Gaussian splatting and a lightweight alternative to diffusion-based inpainting. The paper provides a reproducible GitHub link, extensive ablations, and a clear pipeline. The main contributions—patch-level rasterization and DINO-guided semantic alignment—are plausible and the component-wise ablations show consistent improvements. However, the central evidence for 'semantic consistency' is thin (only mean cosine similarity and t-SNE), and the performance claims are overstated relative to Table 1. The required additional analyses are feasible and would strengthen the paper without changing its core approach.

major comments (4)
  1. [§3.4, Eq. (6)] The semantic alignment loss is evaluated only by mean cosine similarity and t-SNE (Fig. 3). Because Lalign is a negative cosine between fpred and the sample-specific fclean, a representation that collapses toward the mean feature direction can achieve high average cosine without encoding the missing region's content. Since global semantic consistency is a central claim, please provide evidence that fpred is sample-specific: e.g., distribution of pairwise cosine similarities among fpred for different samples, a linear-probe/scene-classification accuracy, or comparison against a mean-feature baseline in the downstream inpainting. The current ablations show the module helps FID/LPIPS but do not establish semantic content.
  2. [§4.3, Table 1] The claim 'matches or outperforms state-of-the-art baselines in both fidelity and perceptual quality' is not supported by Table 1: the method's FID is higher than LaMa in all five settings (e.g., CelebA-HQ small 6.38 vs 5.26; Places2 small 2.61 vs 1.43) and higher than MAT on Places2 large (5.03 vs 3.47). Only LPIPS is consistently better. Please revise the claim to reflect that the method is competitive in LPIPS but not in FID, and report variance/significance (e.g., at least three runs) to support 'competitive' given the small metric gaps.
  3. [§1 / §2.2] The contribution statement 'first approach to leverage 2D Gaussian splatting for high-quality image inpainting' is in tension with the cited work (Fein-Ashley and Fein-Ashley 2024), which also applies Gaussian splatting to inpainting. Please clarify the technical novelty relative to that work (e.g., patch-level rasterization vs diffusion-based use of GS) or soften the priority claim.
  4. [§3.3 / §4.3, Table 4] The patch-level rasterization is motivated by memory reduction, but Table 4 reports only inference time. No peak-memory comparison is given, so the claimed GPU-memory advantage is unverified. Please include memory measurements (e.g., with/without patch strategy) to substantiate this contribution.
minor comments (4)
  1. [§4.4 heading] 'Abalation' should be 'Ablation'.
  2. [Eq. (5)] The symbol x is undefined; the AdaLN operation should be written with the hidden feature map f (or LN(f)).
  3. [§4.4, Figure references] The text says 'more artifacts and structural inconsistencies observed ... as shown in Figure 5'; Figure 5 is the convergence plot, not the visual comparison. Please correct the cross-reference.
  4. [Table 1] For ZITS++, the '-' entries are unclear; indicate whether the model was not applicable or not run on that setting.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark-driven evaluation; DINO alignment module is trained, not a predetermined output.

full rationale

The derivation chain is self-contained. The inpainted image is produced by an end-to-end encoder-rasterization pipeline optimized with reconstruction, adversarial, perceptual, and alignment losses (Eq. 7), and the main claims are supported by external benchmarks (Table 1) and ablations (Table 2) on held-out image-mask pairs against published baselines. The DINO feature adaptation MLP is trained with Lalign = -cos(fclean, fpred); reporting in Figure 3 that cosine similarity improves after training is a diagnostic of that objective rather than an independent prediction, and the paper's semantic-consistency claim does not reduce to this diagnostic: the ablations with and without the DINO map and the full model show downstream FID/LPIPS differences that are not forced by construction. No load-bearing step relies on a self-citation: the Gaussian prior initialization is adopted from external prior work (Peng et al. 2025), and DINOv2, StyleGAN, LaMa, etc. are all external. The Appendix limitation about lacking controllability is an honest statement, not a circular step. The concern that Lalign may permit a generic feature centroid is a legitimate robustness critique but not circularity, because Eq. 6 does not by construction force that collapse; it is an empirical question.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No novel physical entities. The free parameters are conventional hyperparameters and loss weights; the key domain assumption is that masked images suffice to predict complete Gaussian fields, and that DINO features remain adaptable for large masks.

free parameters (6)
  • loss weights (w1:w2:w3:w4) = 1:0.2:2:1 (CelebA-HQ), 1:0.3:3:1 (Places2)
    Empirically set in Eq. 7 and Table 3; not derived.
  • patch size p and overlap pad a = 16x16, pad=1
    Default hyperparameters in Table 3; no ablation on patch size/overlap.
  • Gaussians per patch Npatch = 324 (ablated 100, 196)
    Number of Gaussian kernels per patch; ablation shows quality scales with N.
  • per-Gaussian hidden dimension = 12
    Hidden embedding size per Gaussian element (Table 3).
  • mask ratio thresholds = small 20-40%, large 40-60%
    Definition used to split results in Tables 1-2; chosen by convention, not justified.
  • DINO feature mapping MLP = unspecified
    The lightweight MLP in Section 3.4 has no stated architecture or size, but the Lalign loss depends on its output.
assumptions (5)
  • standard math Image pixels can be represented by a mixture of 2D Gaussians (Eq. 2).
    GMM representation from Zhang et al. 2024 and Reynolds 2009, invoked in Section 3.1.
  • domain assumption A complete Gaussian parameter field can be predicted from a masked image alone, without the binary mask.
    Stated as the core hypothesis in Section 1: 'a complete Gaussian representation of an image can be modeled from its incomplete regions'.
  • domain assumption DINOv2 features extracted from masked images can be transformed to clean features by a lightweight MLP, and these adapted features provide effective global semantic guidance.
    Section 3.4 assumes the mapping block generalizes across mask sizes; supported only by Figure 3 measurements on coarse mask ratios.
  • domain assumption AdaLN modulation can inject semantic conditioning into the U-Net bottlenecks without destabilizing training.
    Section 3.4 and ablation in Table 2; removing gamma causes training collapse after 50k steps.
  • domain assumption Patch overlap blending with weighted averages ensures boundary continuity.
    Section 3.3 describes blending but no ablation validates the overlap strategy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of 2D Gaussian Splatting with Semantic Alignment for Image Inpainting." pith.science (2026). https://pith.science/paper/SCFOTKGR

@misc{pith2026250901964,
  author       = {Pith},
  title        = {Pith review of: 2D Gaussian Splatting with Semantic Alignment for Image Inpainting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SCFOTKGR}},
  note         = {Machine review of arXiv:2509.01964}
}
read the original abstract

Gaussian Splatting (GS), a recent technique for converting discrete points into continuous spatial representations, has shown promising results in 3D scene modeling and 2D image super-resolution. In this paper, we explore its untapped potential for image inpainting, which demands both locally coherent pixel synthesis and globally consistent semantic restoration. We propose the first image inpainting framework based on 2D Gaussian Splatting, which encodes incomplete images into a continuous field of 2D Gaussian splat coefficients and reconstructs the final image via a differentiable rasterization process. The continuous rendering paradigm of GS inherently promotes pixel-level coherence in the inpainted results. To improve efficiency and scalability, we introduce a patch-wise rasterization strategy that reduces memory overhead and accelerates inference. For global semantic consistency, we incorporate features from a pretrained DINO model. We observe that DINO's global features are naturally robust to small missing regions and can be effectively adapted to guide semantic alignment in large-mask scenarios, ensuring that the inpainted content remains contextually consistent with the surrounding scene. Extensive experiments on standard benchmarks demonstrate that our method achieves competitive performance in both quantitative metrics and perceptual quality, establishing a new direction for applying Gaussian Splatting to 2D image processing.

Figures

Figures reproduced from arXiv: 2509.01964 by the authors.

Figure 1
Figure 1. (a) Traditional inpainting methods rely on CNNs [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our proposed framework: (a) The overall pipeline consists of a U-Net architecture, DINO-based semantic [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Representation Gap with Mask Image: (a) Cosine similarity between masked and unmasked DINOv2 features drops [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Qualitative Results. The top two rows show face inpainting results from the CelebA-HQ dataset, while the bottom [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Convergence Speed comparison between CNN de [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Ablation study. Visual comparison of inpainting [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: More results on Real-world Scenarios F Efficiency To assess the computational efficiency of our method, we compare its inference speed with several representative in￾painting approaches, as summarized in [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results on CelebA-HQ [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Qualitative results on Places2 [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Qualitative results on FFHQ [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Qualitative results on ImageNet-100 [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Forecasting as Rendering: A 2D Gaussian Splatting Framework for Time Series Forecasting

    cs.LG 2026-02 conditional novelty 6.0 of 10

    TimeGS forecasts time series by rasterizing learned Gaussian kernels on a period-phase grid, but its state-of-the-art claim is contradicted by its own benchmark table.

Reference graph

Works this paper leans on

53 extracted references · 41 canonical work pages · cited by 1 Pith paper

  1. [1]

    Cao, C.; Dong, Q.; and Fu, Y. 2023. Zits++: Image inpainting by improving the incremental transformer on structural priors. IEEE transactions on pattern analysis and machine intelligence, 45(10): 12667--12684

    work page 2023

  2. [2]

    Cao, J.; Wang, Q.; Xian, Y.; Li, Y.; Ni, B.; Pi, Z.; Zhang, K.; Zhang, Y.; Timofte, R.; and Van Gool, L. 2023. Ciaosr: Continuous implicit attention-in-attention network for arbitrary-scale image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 1796--1807

    work page 2023

  3. [3]

    Chang, H.; Zhang, H.; Jiang, L.; Liu, C.; and Freeman, W. T. 2022. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11315--11325

    work page 2022

  4. [4]

    Chen, D.; Chen, L.; Zhang, Z.; and Zhang, L. 2025. Generalized and Efficient 2D Gaussian Splatting for Arbitrary-scale Super-Resolution. arXiv preprint arXiv:2501.06838

    work page Pith review arXiv 2025

  5. [5]

    Chen, H.; and Zhao, Y. 2024. Don't Look into the Dark: Latent Codes for Pluralistic Image Inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7591--7600

    work page 2024

  6. [6]

    Chen, Y.; Liu, S.; and Wang, X. 2021. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 8628--8638

    work page 2021

  7. [7]

    Chen, Z.; Wang, F.; Wang, Y.; and Liu, H. 2024. Text-to-3d using gaussian splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 21401--21412

    work page 2024

  8. [8]

    Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei-Fei, L. 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, 248--255. Ieee

    2009

Show all 53 references
  1. [9]

    Deng, J.; Wu, X.; Yang, Y.; Zhu, C.; Wang, S.; and Wu, Z. 2025. Acquire and then Adapt: Squeezing out Text-to-Image Model for Image Restoration. arXiv preprint arXiv:2504.15159

    2025 arXiv

  2. [10]

    Deng, Y.; Hui, S.; Zhou, S.; Meng, D.; and Wang, J. 2022. T-former: An efficient transformer for image inpainting. In Proceedings of the 30th ACM international conference on multimedia, 6559--6568

    2022

  3. [11]

    Dong, J.; Wang, C.; Zheng, W.; Chen, L.; Lu, J.; and Tang, Y. 2025. GaussianToken: An Effective Image Tokenizer with 2D Gaussian Splatting. arXiv preprint arXiv:2501.15619

    2025 arXiv

  4. [12]

    A.; and Leung, T

    Efros, A. A.; and Leung, T. K. 1999. Texture synthesis by non-parametric sampling. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, 1033--1038. IEEE

    1999

  5. [13]

    Esser, P.; Rombach, R.; and Ommer, B. 2021. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 12873--12883

    2021

  6. [14]

    Fein-Ashley, J.; and Fein-Ashley, B. 2024. Diffusion Models with Anisotropic Gaussian Splatting for Image Inpainting. arXiv:2412.01682

    2024 arXiv

  7. [15]

    J.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; and Bengio, Y

    Goodfellow, I. J.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; and Bengio, Y. 2014. Generative adversarial nets. Advances in neural information processing systems, 27

    2014

  8. [16]

    Guo, X.; Yang, H.; and Huang, D. 2021. Image inpainting via conditional texture and structure dual generation. In Proceedings of the IEEE/CVF international conference on computer vision, 14134--14143

    2021

  9. [17]

    He, K.; and Sun, J. 2014. Image completion approaches using the statistics of similar patches. IEEE transactions on pattern analysis and machine intelligence, 36(12): 2423--2435

    2014

  10. [18]

    Heusel, M.; Ramsauer, H.; Unterthiner, T.; Nessler, B.; and Hochreiter, S. 2017. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30

    2017

  11. [19]

    Hu, J.; Xia, B.; Chen, B.; Yang, W.; and Zhang, L. 2025. Gaussiansr: High fidelity 2d gaussian splatting for arbitrary-scale image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 3554--3562

    2025

  12. [20]

    Huang, Y.-H.; Sun, Y.-T.; Yang, Z.; Lyu, X.; Cao, Y.-P.; and Qi, X. 2024. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 4220--4230

    2024

  13. [21]

    Jiang, C.; Jia, H.; Xu, H.; Ye, W.; Dong, M.; Yan, M.; Zhang, J.; Huang, F.; and Zhang, S. 2024. Maven: An effective multi-granularity hybrid visual encoding framework for multimodal large language model. Advances in Neural Information Processing Systems, 37: 101992--102010

    2024

  14. [22]

    Jiao, Y.; Qiu, H.; Jie, Z.; Chen, S.; Chen, J.; Ma, L.; and Jiang, Y.-G. 2025. Unitoken: Harmonizing multimodal understanding and generation through unified visual encoding. arXiv preprint arXiv:2504.04423

    2025 arXiv

  15. [23]

    Johnson, J.; Alahi, A.; and Fei-Fei, L. 2016. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, 694--711. Springer

    2016

  16. [24]

    Ju, X.; Liu, X.; Wang, X.; Bian, Y.; Shan, Y.; and Xu, Q. 2024. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In European Conference on Computer Vision, 150--168. Springer

    2024

  17. [25]

    Karras, T.; Aila, T.; Laine, S.; and Lehtinen, J. 2017. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196

    2017 arXiv

  18. [26]

    Karras, T.; Laine, S.; and Aila, T. 2019. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 4401--4410

    2019

  19. [27]

    M.; Yang, G.; Scherer, S.; Ramanan, D.; and Luiten, J

    Keetha, N.; Karhade, J.; Jatavallabhula, K. M.; Yang, G.; Scherer, S.; Ramanan, D.; and Luiten, J. 2024. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21357--21366

    2024

  20. [28]

    Kerbl, B.; Kopanas, G.; Leimk \"u hler, T.; and Drettakis, G. 2023. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4): 139--1

    2023

  21. [29]

    Li, H.; Dai, T.; Li, Y.; Zou, X.; and Xia, S.-T. 2022 a . Adaptive local implicit image function for arbitrary-scale super-resolution. In 2022 IEEE International Conference on Image Processing (ICIP), 4033--4037. IEEE

    2022

  22. [30]

    Li, W.; Lin, Z.; Zhou, K.; Qi, L.; Wang, Y.; and Jia, J. 2022 b . Mat: Mask-aware transformer for large hole image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10758--10768

    2022

  23. [31]

    A.; Shih, K

    Liu, G.; Reda, F. A.; Shih, K. J.; Wang, T.-C.; Tao, A.; and Catanzaro, B. 2018. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European conference on computer vision (ECCV), 85--100

    2018

  24. [32]

    Liu, W.; Cun, X.; Pun, C.-M.; Xia, M.; Zhang, Y.; and Wang, J. 2023. Coordfill: Efficient high-resolution image inpainting via parameterized coordinate querying. In Proceedings of the AAAI conference on artificial intelligence, volume 37, 1746--1754

    2023

  25. [33]

    Lugmayr, A.; Danelljan, M.; Romero, A.; Yu, F.; Timofte, R.; and Van Gool, L. 2022. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11461--11471

    2022

  26. [34]

    Luiten, J.; Kopanas, G.; Leibe, B.; and Ramanan, D. 2024. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 2024 International Conference on 3D Vision (3DV), 800--809. IEEE

    2024

  27. [35]

    P.; Tancik, M.; Barron, J

    Mildenhall, B.; Srinivasan, P. P.; Tancik, M.; Barron, J. T.; Ramamoorthi, R.; and Ng, R. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1): 99--106

    2021

  28. [36]

    Oquab, M.; Darcet, T.; Moutakanni, T.; Vo, H. V.; Szafraniec, M.; Khalidov, V.; Fernandez, P.; Haziza, D.; Massa, F.; El-Nouby, A.; Howes, R.; Huang, P.-Y.; Xu, H.; Sharma, V.; Li, S.-W.; Galuba, W.; Rabbat, M.; Assran, M.; Ballas, N.; Synnaeve, G.; Misra, I.; Jegou, H.; Maira...

    2023

  29. [37]

    Pathak, D.; Krahenbuhl, P.; Donahue, J.; Darrell, T.; and Efros, A. A. 2016. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2536--2544

    2016

  30. [38]

    Peng, L.; Wu, A.; Li, W.; Xia, P.; Dai, X.; Zhang, X.; Di, X.; Sun, H.; Pei, R.; Wang, Y.; et al. 2025. Pixel to gaussian: Ultra-fast continuous super-resolution with 2d gaussian modeling. arXiv preprint arXiv:2503.06617

    2025 arXiv

  31. [39]

    A.; et al

    Reynolds, D. A.; et al. 2009. Gaussian mixture models. Encyclopedia of biometrics, 741(659-663): 3

    2009

  32. [40]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10684--10695

    2022

  33. [41]

    Suvorov, R.; Logacheva, E.; Mashikhin, A.; Remizova, A.; Ashukha, A.; Silvestrov, A.; Kong, N.; Goka, H.; Park, K.; and Lempitsky, V. 2022. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of co...

    2022

  34. [42]

    Tian, C.; Tao, C.; Dai, J.; Li, H.; Li, Z.; Lu, L.; Wang, X.; Li, H.; Huang, G.; and Zhu, X. 2023. Addp: Learning general representations for image recognition and generation with alternating denoising diffusion process. arXiv preprint arXiv:2306.05423

    2023 arXiv

  35. [43]

    Tianyidan, X.; Ma, R.; Wang, Q.; Ye, X.; Liu, F.; Tai, Y.; Zhang, Z.; Wang, L.; and Yi, Z. 2025. Anywhere: A Multi-Agent Framework for User-Guided, Reliable, and Diverse Foreground-Conditioned Image Generation. In Proceedings of the AAAI Conference on Artificial Intelligence, ...

    2025

  36. [44]

    Wasserman, N.; Rotstein, N.; Ganz, R.; and Kimmel, R. 2024. Paint by inpaint: Learning to add image objects by removing them first. arXiv preprint arXiv:2404.18212

    2024 arXiv

  37. [45]

    Wu, H.; Ni, N.; and Zhang, L. 2023. Learning dynamic scale awareness and global implicit functions for continuous-scale super-resolution of remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 61: 1--15

    2023

  38. [46]

    Xie, S.; Zhang, Z.; Lin, Z.; Hinz, T.; and Zhang, K. 2023. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 22428--22437

    2023

  39. [47]

    Yu, J.; Lin, Z.; Yang, J.; Shen, X.; Lu, X.; and Huang, T. S. 2018. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, 5505--5514

    2018

  40. [48]

    Yu, J.; Lin, Z.; Yang, J.; Shen, X.; Lu, X.; and Huang, T. S. 2019. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, 4471--4480

    2019

  41. [49]

    Yu, T.; Feng, R.; Feng, R.; Liu, J.; Jin, X.; Zeng, W.; and Chen, Z. 2023. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790

    2023 arXiv

  42. [50]

    A.; Shechtman, E.; and Wang, O

    Zhang, R.; Isola, P.; Efros, A. A.; Shechtman, E.; and Wang, O. 2018. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, 586--595

    2018

  43. [51]

    Zhang, X.; Ge, X.; Xu, T.; He, D.; Wang, Y.; Qin, H.; Lu, G.; Geng, J.; and Zhang, J. 2024. Gaussianimage: 1000 fps image representation and compression by 2d gaussian splatting. In European Conference on Computer Vision, 327--345. Springer

    2024

  44. [52]

    Zheng, C.; Cham, T.-J.; and Cai, J. 2019. Pluralistic image completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1438--1447

    2019

  45. [53]

    Zhou, B.; Lapedriza, A.; Khosla, A.; Oliva, A.; and Torralba, A. 2017. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6): 1452--1464

    2017

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.