Pith. sign in

REVIEW 4 major objections 7 minor 37 references

GEAR: Reconstruction of Classical Paintings via Geometry Grounding and Appearance Restitution

T0 review · 4 major / 7 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A classical painting can become a plausible 3D scene by splitting the task: first convert the flat artwork into a sculpture-like image with coherent shading, reconstruct 3D Gaussians from it, then restore the original brushwork under multi-

desk verdict A useful new task and a sensible two-stage recipe, but the quantitative evidence for the geometry claim does not hold up yet. read the letter →

arxiv 2607.17519 v2 pith:TR4AFMD4 submitted 2026-07-20 cs.MM

Qinyu Zhang , Xinda Liu , Yunchen Li , Yunzhuo Liu , Chenxi Hu , Kang Li , Guohua Geng This is my paper · ORCID
classification cs.MM
keywords 3DreconstructionfromsingleimageclassicalpaintingsGaussiansplattingculturalheritageappearancefidelitygeometrygroundingnovelviewsynthesisbenchmarkdataset
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 central claim is that a single classical painting — despite stylized perspective, flat shading, and ambiguous depth — can be turned into a plausible, explorable 3D scene if the task is split into two stages rather than forced through one representation. In the first stage, the painting is translated into a sculpture-like version with more coherent lighting and shading cues, from which a pretrained single-view 3D reconstruction model produces stable geometry. In the second stage, that grounded scene is edited through diffusion-based multi-view appearance targets, restoring painterly texture and detail while an anchor constraint keeps the geometry from drifting. If the claim holds, the practical consequence is a training-free recipe for cultural-heritage digitization: artworks that currently defeat natural-image reconstruction pipelines can be reconstructed without retraining, and the same two-stage separation may transfer to other stylized image domains. The authors support the claim with a new benchmark of 10,160 artworks and with geometry metrics, image-quality-judge evaluations, and a user study.

What carries the argument

The load-bearing mechanism is the two-stage separation itself, with the adaptive illumination-grounding operation as the concrete engine of the first stage. The method computes a log-ratio illumination field R = log(L+ε) − log(gmean(L+ε)) from the translated image, smooths it locally and at multiple scales with boundary-aware weights, then fuses the smooth and multi-scale fields with a contrast-adaptive coefficient α_ms = clip(α_base · α_target/(σ_R + ε), α_min, α_max). Recombining this grounded illumination with the original painting's reflectance (and re-adding high-frequency brushwork detail) yields an input whose shading cues are more consistent for a pretrained single-view reconstructio

What would settle it

Take a set of classical paintings whose depicted scenes have known 3D structure (for instance, paintings made from a real scene or a 3D model, or a 3D scene rendered in painterly style). Compare the depth maps recovered by GeAR from the painting with depth recovered from the original photograph. If grounding does not reduce depth error relative to direct reconstruction from the painting, the central claim is falsified.

Watch

Extended reading notes

Core claim

The discovery on the paper's own terms is that geometric recoverability and painterly fidelity are not conflicting demands on a single representation but separable stages. Geometry Grounding maps the painting to a geometry-oriented image using an edge-conditioned style translation (with a 'sculpture' prompt) plus an illumination-grounding step: it computes a log-ratio illumination field, smooths it with boundary-aware multi-scale filtering, adaptively weights the smoothing by global contrast, and recombines it with the original reflectance and high-frequency details. Feeding this grounded image to a pretrained monocular Gaussian reconstruction model stabilizes the recovered structure. Appear

Load-bearing premise

The claim rests on the premise that a text-to-image model prompted with 'sculpture' turns a painting into an image whose shading and illumination are physically coherent enough that the geometry inferred from it is genuinely better, not just higher on shape-variation scores.

Editorial extensions

If this is right

  • If GeAR is right, classical paintings that currently produce flat, unstable reconstructions from single-view models can be reconstructed without any training or fine-tuning, using only prompt engineering, illumination grounding, and a generic editing model.
  • The two-stage recipe should preserve geometry gains through the appearance stage: the paper's ablation shows geometric metrics rise after Appearance Restitution, not fall.
  • The method applies across a wide span of artistic traditions — murals, ukiyo-e prints, Tibetan thangkas, Persian miniatures, court-lady paintings — suggesting the approach is not tied to one painting style.
  • Grounding should outperform generic preprocessing (histogram equalization, retinex) because it specifically constructs geometry-compatible shading rather than merely enhancing contrast.
  • Appearance fidelity and geometric plausibility can be measured separately and both improved, which gives the new task a concrete evaluation protocol for future work.

Reading between the lines

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

  • A sharp reader should treat the sculpture-prompt step as an empirical bet: the claim that the translation carries 'physically consistent lighting' is asserted, not verified against geometric ground truth. A direct test — reconstructing from a photograph-like rendering of the same painted scene and comparing depth error — would settle whether grounding improves real geometry or just shape-diversity
  • The two-stage separation could transfer beyond paintings: any stylized single image (caricature, anime cel, stained-glass window) with weak photometric cues might benefit from the same grounding-then-restitution split, provided a 'grounding' prompt can be found that preserves layout while adding shading coherence.
  • The benchmark's lack of 3D ground truth means current metrics measure plausibility, not accuracy. If future work adds a small set of scenes with known geometry (e.g., a 3D scene rendered in painterly style), the whole evaluation could be anchored to actual correspondence rather than relative preference.
  • Releasing the dataset and code could let the community stress-test whether the method's gains hold under different diffusion-editing backbones and different single-view reconstruction models, since the pipeline is training-free and modular.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper introduces Classical Painting-to-3D (CP3D), a task and benchmark for reconstructing 3D Gaussian scenes from a single classical painting while jointly maintaining geometric plausibility and fidelity to the source artwork. The proposed method, GeAR, is a training-free two-stage pipeline: Geometry Grounding first translates the painting into a more geometry-oriented image via edge-guided style transfer and illumination-ratio manipulation (Eqs. 2–13), then reconstructs a Gaussian scene with TRELLIS; Appearance Restitution then edits multi-view renders with a diffusion model and refines the Gaussians with anchor-constrained optimization (Eqs. 14–16). The authors build HeriArch, a 10,160-image benchmark, and report quantitative gains in VEND/DDI geometry metrics, MLLM appearance scores, and a 300-participant user study over SplatterImage, LGM, and TRELLIS.

Significance. If the claims were fully supported, the main contribution would be an effective inference-time recipe for a genuinely hard domain: classical paintings violate natural-image assumptions, and decoupling geometric recoverability from painterly appearance is a plausible design solution. HeriArch, with 10,160 images across six artistic traditions, is a potentially valuable resource, and the user study (300+ participants, including experts) is a useful qualitative complement. The comparison against generic enhancement baselines (Table 6) is also a good sanity check. However, the central quantitative evidence for geometric improvement is currently unsupported: the novelty metrics are unvalidated, the stage ablation is consistent with metric contamination by texture, and the key prompt is selected on the evaluation benchmark. The significance therefore remains conditional until these concerns are addressed.

major comments (4)
  1. [§5.2, Table 1; §5.1, Eq. (16), Table 3] The central claim of geometric plausibility rests entirely on VEND and DDI, whose definitions are deferred to an appendix and which are never validated against ground-truth geometry. Both definitions are of the form that can be inflated by non-geometric signal: VEND rewards volumetric extent and normal directional diversity, and DDI rewards high-frequency geometric variation. Table 3 makes the risk concrete: Appearance Restitution, an appearance-only refinement stage whose optimization is anchored to the grounded geometry by L_anchor (Eq. 16), increases DDI_avg from 6.68 to 10.24 (+53%) and VEND from 85.84 to 87.61. If geometry is held approximately constant by the anchor term, a depth-detail metric should not nearly double from texture/color editing. The authors should validate VEND/DDI on scenes with known 3D geometry (e.g., synthetic renders of 3D models with and without added texture
  2. [§5.6, Table 4; Eq. (3); §5.2, Table 1] The Sculpture prompt T_g is a design choice that is selected by comparing prompt styles on the HeriArch benchmark and taking the one that maximizes VEND/DDI (Table 4: Sculpture 11.05 DDI vs 2.46 for Pencil sketch). The same benchmark is then used in Table 1 to claim that GeAR outperforms the baselines. This is a selection-on-the-test-set problem and can materially inflate the reported advantage. The prompt should be fixed a priori on a validation split, or the final comparison should be on a held-out set that was not used for any prompt/ablation choice. Without this, the 'consistent outperformance' claim is not properly supported.
  3. [Tables 1, 3, 5] No variance or confidence interval is reported for any geometry metric, even though Table 3 reports N=1000. Several differences in Table 5 (e.g., VEND 85.29 vs 85.27 vs 86.01) are much smaller than plausible sample-to-sample noise at that scale. The paper should report means with standard deviations or bootstrap CIs over at least several independent runs (or resampling of the benchmark), and state the number of runs. Without this, the reader cannot tell whether the reported gains are significant.
  4. [§5.1] The availability of the appendix is essential: the full text states that 'Detailed definitions of VEND, DDI, and the full evaluation protocols are provided in the appendix,' but the appendix is not included in the reviewed version. Because the main quantitative results depend on these definitions, the submission is not currently reproducible. Please include the appendix (or define the metrics in the main text) in the revision.
minor comments (7)
  1. [Title/Abstract] The model is called GeAR throughout the text but GEAR in the title; unify.
  2. [Fig. 4 caption] The caption contains a formatting artifact: 'Splatter [26Image ] LGM'.
  3. [Eq. (4) vs Eq. (14)] R denotes both the log-ratio illumination field and the rendered views; use different symbols to avoid confusion.
  4. [§5.4] The user study should report the raw counts behind Fig. 3(b) and confidence intervals for the preference percentages; the expertise distribution in Fig. 3(a) is informative but not sufficient.
  5. [Eq. (16)] L_edit and L_anchor are not defined in the main text, and lambda_anc is deferred to the appendix; the losses should be specified.
  6. [Table 3] The use of NIQE for stylized painterly content, although described as a supplementary diagnostic, deserves one sentence of justification since NIQE was not designed for non-photographic images.
  7. [Abstract/Conclusion] The paper promises that code and dataset will be released; please include a working link or anonymized repository in the revision to support reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline geometry gains are partly fitted to the benchmark via prompt selection; DDI behavior suggests the metric rewards appearance, not geometry.

  1. fitted input called prediction [Eq. (3); Sec. 5.6 Table 4; Sec. 5.2 Table 1]
    "Prompt Style VEND↑ DDI avg↑ NIQE avg↓ / Sculpture 85.58 11.05 4.79 / Pencil sketch 83.86 2.46 7.25 / Woodcut 80.46 2.95 6.79 ... the sculpture-style prompt significantly improves both VEND and DDI scores"

    The geometry-grounding prompt T_g in Eq. (3) is an input to GeAR. Table 4 explicitly searches over prompt styles, choosing 'Sculpture' because it maximizes the same VEND/DDI metrics used in Table 1 to claim GeAR outperforms TRELLIS. Thus the headline geometric improvement is partly the selection criterion itself, not an independent prediction: with a different fixed prompt the reported DDI drops from 11.05 to 2.46 or 2.95. The method's internal hyperparameter is fitted to the evaluation benchmark and then presented as a method-level result.

  2. other [Secs. 3.2.2 and 5.6; Eq. (16); Table 3]
    "GeAR w/o AR ... DDI avg 6.68; GeAR w/ AR ... DDI avg 10.24 ... adding Appearance Restitution improves both VEND and DDI, indicating that the second stage preserves and further enhances geometric plausibility after Geometry Grounding."

    Appearance Restitution, Eq. (16), is an appearance-only refinement: it optimizes toward diffusion-edited appearance targets while the anchor term is intended to 'suppress geometric drift.' A 53% DDI increase from such an appearance-only stage is inconsistent with DDI measuring depth detail integrity; it indicates that the metric is responding to added high-frequency texture/opacity or shading variation rather than to actual depth structure. The paper's central claim of geometric improvement therefore rests on a metric whose behavior is not geometry-specific, and the 'geometric plausibility' conclusion reduces to a metric artifact unless the deferred appendix definition shows otherwise.

full rationale

GeAR's derivation is not literally Eq. X = Eq. Y, and there is no load-bearing self-citation or imported uniqueness theorem: TRELLIS and GaussianEditor are external tools, and the user study provides some independent evidence. However, the main quantitative geometry claim is partially circular. The 'Sculpture' prompt, a core input of Geometry Grounding, is selected by inspecting VEND/DDI on the same HeriArch benchmark on which GeAR is then declared superior; the benchmark comparison is therefore not a fresh evaluation of a fixed, pre-specified pipeline. Moreover, the internal ablation shows that an appearance-only refinement stage raises DDI from 6.68 to 10.24, strongly suggesting the geometry metric is sensitive to appearance variation, while the manuscript explicitly defers the definitions ('Detailed definitions of VEND, DDI, and the full evaluation protocols are provided in the appendix'), so the claimed geometric reduction cannot be verified from the available text. The human-preference and MLLM appearance results give the work independent content, so this is partial circularity/selection bias rather than full definitional circularity.

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

No new physical or geometric entity is proposed; the ledger is dominated by free parameters chosen on the paper's own benchmark (prompt style, fusion ranges, anchor weight are either swept or hidden in the appendix) and ad hoc evaluation axioms (VEND/DDI as truth). This is the main epistemic weakness of the paper: the benchmark is simultaneously the tuning set, the demonstration set, and the source of the claimed superiority.

free parameters (4)
  • Sculpture translation prompt Tg = "Sculpture" (selected over "Pencil sketch" and "Woodcut" in Table 4)
    The prompt is the central knob of Geometry Grounding; Table 4 is an explicit ablation selecting the prompt that maximizes VEND/DDI on HeriArch. Choosing it on the same benchmark used for final claims inflates the reported gains.
  • alpha_base / alpha_target / alpha_min / alpha_max = 0.8 / 1.0 / 0.4 / 1.5
    Adaptive fusion-coefficient settings in Eq. (7)–(8); Table 5 sweeps the range to select [0.4,1.5] as "the best overall balance," i.e., tuned on HeriArch.
  • beta (illumination grounding strength) = not reported in main text
    Strength of the illumination-injection term in Eq. (10); no ablation or value is given in the main text, so it is an unexamined hand-set constant.
  • anchor weight lambda_anc and 7 rendered viewpoints = n=7; lambda_anc deferred to appendix
    Eq. (16) balances appearance-editing loss against geometry preservation; the weight is unstated, and the number of viewpoints is chosen without ablation in the main text.
assumptions (4)
  • domain assumption A text-guided image translation model conditioned on an edge map preserves scene layout sufficiently for monocular reconstruction (Eq. 3).
    Load-bearing: if translation distorts layout, the grounded geometry inherits the distortion. The paper tests only downstream metrics, not layout preservation itself.
  • ad hoc to paper VEND and DDI are meaningful proxies for "geometric plausibility" in the absence of any 3D ground truth.
    These metrics are introduced in this paper and defined only in an appendix we do not have; they measure volumetric extent, normal diversity, and depth-variation-like quantities, not correspondence to real scene geometry.
  • domain assumption The MLLM judge's four-dimension scores (TF, GA, VC, AC) reflect artwork fidelity and perceptual quality on blind unpaired outputs.
    MLLM-as-a-judge is applied to textured 3D renderings without human-calibrated thresholds or a published prompt template; the cited "LLM-as-a-judge" literature itself documents position and verbosity biases.
  • domain assumption TRELLIS is a suitable pretrained monocular reconstruction backbone for stylized painting content after grounding.
    TRELLIS is trained on natural/shaded object data; whether its geometry prior is valid for painterly subject matter is tested only through proxy metrics, not true geometry.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GEAR: Reconstruction of Classical Paintings via Geometry Grounding and Appearance Restitution." pith.science (2026). https://pith.science/paper/TR4AFMD4

@misc{pith2026260717519,
  author       = {Pith},
  title        = {Pith review of: GEAR: Reconstruction of Classical Paintings via Geometry Grounding and Appearance Restitution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TR4AFMD4}},
  note         = {Machine review of arXiv:2607.17519}
}
read the original abstract

Classical paintings preserve rich spatial, cultural, and historical content, making their reconstruction as explorable 3D scenes valuable for digital preservation, immersive exhibition, and cultural engagement. Yet, unlike photographs, they often depict scenes in a single-view, stylized manner, with weak perspective, lighting, and depth cues. Existing 3D reconstruction methods are largely built on natural-image priors, making it difficult to recover geometrically plausible and visually faithful 3D representations from such inputs. To address this challenge, we introduce Classical Painting-to-3D (CP3D), a new task that aims to recover a 3D representation from a single classical painting while jointly ensuring geometric plausibility, appearance fidelity to the source artwork, and plausible novel-view synthesis. We further propose GeAR, a training-free two-stage framework for Geometry Grounding and Appearance Restitution. GeAR first converts the input painting into a geometry-grounded representation with more coherent shading and illumination cues, improving the stability of 3D Gaussian reconstruction. It then restores artwork-faithful appearance across views under spatial constraints and multi-view consistency, recovering the painterly textures and details weakened during grounding. In addition, we construct HeriArch, a curated benchmark of 10,160 high-resolution classical artworks for systematic evaluation of CP3D. Extensive experiments and user studies show that GeAR consistently outperforms strong baselines in geometric plausibility, appearance fidelity, and human preference. Code and dataset will be released publicly.

Figures

Figures reproduced from arXiv: 2607.17519 by the authors.

Figure 1
Figure 1. Motivation of Classical Painting-to-3D (CP3D) and the domain gap between classical paintings and natural images. (a) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of GeAR. We first perform Geometry Grounding to transform a classical painting into a geometry-oriented [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) Distribution of participants across three exper [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparisons of generated 3D Gaussians for image-to-3D. Including Mural, Indian Art, Persian Miniature, Court [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Qualitative Comparison of Reconstruction Results [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 2 linked inside Pith

  1. [1]

    Titas Anciukevičius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. 2023. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 12608–12618

    2023

  2. [2]

    Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, Sumith Kulal, et al. 2025. Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space.arXiv e-prints(2025), arXiv–2506

    2025

  3. [3]

    David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann

  4. [4]

    Ruiyang Chen, Mohan Yin, Jiawei Shen, and Wei Ma. 2024. Recon3D: High Quality 3D Reconstruction from a Single Image Using Generated Back-View Explicit Priors. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2802–2811

    2024

  5. [5]

    Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. 2024. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 21476–21485

    2024

  6. [6]

    Andreas Dahn, Leszek Plichta, Simon Spielmann, Eduard Schäfer, and Justus Blönnigen. 2024. Fate of the Minotaur: A scalable location based VR experience. InACM SIGGRAPH 2024 Immersive Pavilion. 1–2

    2024

  7. [7]

    Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. 2024. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.203092, 3 (2024), 4

    arXiv 2024

  8. [8]

    I Ho, Jie Song, Otmar Hilliges, et al . 2024. Sith: Single-view textured human reconstruction with image-conditioned diffusion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 538–549

    2024

Show all 37 references
  1. [9]

    Zixuan Huang, Mark Boss, Aaryaman Vasishta, James M Rehg, and Varun Jam- pani. 2025. Spar3d: Stable point-aware reconstruction of 3d objects from single images. InProceedings of the Computer Vision and Pattern Recognition Conference. 16860–16870

    2025

  2. [10]

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis

  3. [11]

    Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. 2024. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. InEuropean Conference on Computer Vision. Springer, 112–130

    2024

  4. [12]

    Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhattacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, et al

  5. [13]

    Haolong Li, Chenghao Du, Ziheng Jiang, Yifan Zhang, Jiawei Ma, and Chen Ye

  6. [14]

    Luxi Li, Qin Zou, Fan Zhang, Hongkai Yu, Long Chen, Chengfang Song, Xianfeng Huang, and Xiaoguang Wang. 2022. Line Drawing Guided Progressive Inpainting of Mural Damages.ArXiv 2211.06649(2022), 1–12

    2022 arXiv

  7. [15]

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis.Commun. ACM65, 1 (2021), 99–106

    2021

  8. [16]

    Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. 2023. Diffrf: Rendering-guided 3d radi- ance field diffusion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 4328–4338

    2023

  9. [17]

    InProceedings of the AAAI Conference on Artificial Intelligence, Vol

    Towards automated Chinese ancient character restoration: a diffusion- based method with a new dataset. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 3073–3081

  10. [18]

    Qiuhong Shen, Zike Wu, Xuanyu Yi, Pan Zhou, Hanwang Zhang, Shuicheng Yan, and Xinchao Wang. 2025. Gamba: Marry Gaussian Splatting With Mamba for Single-View 3D Reconstruction.IEEE Transactions on Pattern Analysis and Machine Intelligence(2025), 1–14

    2025

  11. [19]

    J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 2023. 3d neural field generation using triplane diffusion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 20875– 20886

    2023

  12. [20]

    Jocelyn Penny Small. 2019. Circling round Vitruvius, linear perspective, and the design of Roman wall painting. InArts, Vol. 8. MDPI, 118

    2019

  13. [21]

    Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. 2022. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition...

    2022

  14. [22]

    Huang SongSheng, Azian Tahir, and Issarezal Ismail. 2025. Re-examining the Meaning of Flatness Based on the Theory Evolution of Medium Specificity in Painting.Asian Journal of Research in Education and Social Sciences7, 1 (2025), 93–104

    2025

  15. [23]

    Matthias Springstein, Stefanie Schneider, Javad Rahnama, Julian Stalter, Maxi- milian Kristen, Eric Müller-Budack, and Ralph Ewerth. 2024. Visual narratives: Large-scale hierarchical classification of art-historical images. InProceedings of the IEEE/CVF Winter Conference on Ap...

    2024

  16. [24]

    Mingze Sun, Junhao Chen, Junting Dong, Yurun Chen, Xinyu Jiang, Shiwei Mao, Puhua Jiang, Jingbo Wang, Bo Dai, and Ruqi Huang. 2025. Drive: Diffusion- based rigging empowers generation of versatile and expressive characters. In Proceedings of the Computer Vision and Pattern Rec...

    2025

  17. [25]

    Nagabhushan Somraj and Rajiv Soundararajan. 2023. Vip-nerf: Visibility prior for sparse input neural radiance fields. InACM SIGGRAPH 2023 conference proceedings. 1–11

    2023

  18. [26]

    Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. 2024. Splatter image: Ultra-fast single-view 3d reconstruction. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10208–10217

    2024

  19. [27]

    Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. 2024. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. InEuropean Conference on Computer Vision. Springer, 1–18

    2024

  20. [28]

    Hanyang Wang, Fangfu Liu, Jiawei Chi, and Yueqi Duan. 2025. VideoScene: Distilling video diffusion model to generate 3D scenes in one step. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 16475– 16485

    2025

  21. [29]

    Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, Joao F Henriques, Christian Rupprecht, and Andrea Vedaldi. 2025. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image. In2025 International Conference on 3D Vision (3DV). IEEE...

    2025

  22. [30]

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

    2025

  23. [31]

    Zishan Xu, Yuqing Yang, Qianzhen Fang, Wei Chen, Tingting Xu, Jueting Liu, and Zehua Wang. 2024. A comprehensive dataset for digital restoration of Dunhuang murals.Scientific Data11, 1 (2024), 955

    2024

  24. [32]

    Jiawei Yang, Marco Pavone, and Yue Wang. 2023. Freenerf: Improving few- shot neural rendering with free frequency regularization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 8254–8263

    2023

  25. [33]

    Han-Lei Wang, Ping-Hsuan Han, Yu-Mu Chen, Kuan-Wen Chen, XinYi Lin, Ming-Sui Lee, and Yi-Ping Hung. 2018. Dunhuang mural restoration using deep learning. InSIGGRAPH Asia 2018 technical briefs. 1–4

    2018

  26. [37]

    Shuo Zhao, Yifei Huang, Xiaoyang He, Xin Tong, Xin Li, and Dan Wu. 2025. Reviving Mural Art through Generative AI: A Comparative Study of AI-Generated and Hand-Crafted Recreations. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–20

    2025

  27. [2023]

    Graph.42, 4 (2023), 139–1

    3D Gaussian splatting for real-time radiance field rendering.ACM Trans. Graph.42, 4 (2023), 139–1

    2023

  28. [2024]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 19457–19467

  29. [2025]

    InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing

    From generation to judgment: Opportunities and challenges of llm-as- a-judge. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2757–2791

    2025

Pith tools

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