Pith. sign in

REVIEW 3 major objections 6 minor 57 references

Fancy123: One Image to High-Quality 3D Mesh Generation via Plug-and-Play Deformation

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Fancy123 claims to make single-image-to-3D meshes sharper by deforming 2D views and the 3D mesh, then unprojecting the original image onto the result.

desk verdict Honest engineering paper whose appearance-enhancement module is genuinely useful, but whose fidelity module has a documented identifiability failure that undercuts the SoTA headline. read the letter →

arxiv 2411.16185 v2 pith:BP45TOHG submitted 2024-11-25 cs.CV

Qiao Yu , Xianzhi Li , Yuan Tang , Xu Han , Long Hu , Yixue Hao , Min Chen This is my paper · ORCID
classification cs.CV
keywords single-image3DreconstructionmultiviewdiffusionmeshdeformationJacobianfieldunprojectioninference-timerefinementtextureclarityappearanceenhancement
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 argues that three problems plaguing one-image-to-3D generation, namely multiview images that do not align, meshes that do not match the input photo, and blurry colors, can be fixed at inference time without retraining. Fancy123 deforms the 2D multiview images to realign them, deforms the 3D mesh with a Jacobian field to match the input, and then unprojects both the input image and the deformed multiview images onto the mesh. The unprojection step deliberately discards the colors predicted by the reconstruction model in favor of the sharp original pixels. A sympathetic reader should care because the two enhancement modules are plug-and-play, meaning they can be dropped into existing pipelines such as InstantMesh, Unique3D, LGM, and others. The paper reports lower FID and LPIPS values and higher PSNR, SSIM, and CLIP-similarity scores than the baselines, while cautioning that the standard metrics often disagree with human perception.

What carries the argument

The machinery is three operations connected in optimization loops: a 2D grid deformation field $F$ that repositions pixels of each multiview image by interpolated grid-vertex offsets; a Jacobian field $J$ for 3D mesh deformation, where per-face Jacobians are optimized and the deformed vertex positions are recovered by solving the Poisson equation $L V' = \nabla^T A J$ with the mesh Laplacian $L$; and unprojection, which colors each mesh vertex as a weighted sum of the pixels that see it, weighted by the cosine similarity between the vertex normal and the camera direction. Differentiable rendering and losses (MSE, mask, and smoothness terms) drive the two enhancement loops. The camera pose of the input image is found by an LPIPS-based coarse-to-fine search so that the fidelity loop has a well-aligned starting point.

What would settle it

Take an input view with two same-colored parts at different depths, such as the paper's bird-and-branch case, run only the fidelity module, and check whether the deformed mesh moves the correct semantic part or instead raises the wrong one; the module's claim fails if the shape change is driven by color matching rather than by geometry.

Watch

Extended reading notes

Core claim

Fancy123's central claim is that explicit deformation at two levels, combined with unprojection, yields state-of-the-art single-image-to-3D mesh quality. Given a multiview diffusion model and a large reconstruction model that produce the initial mesh, the appearance module optimizes a per-view grid of 2D offsets so that the deformed multiview images, when unprojected onto the mesh and re-rendered, match themselves, thereby removing ghosting. The fidelity module then optimizes a Jacobian field so that the mesh, rendered from an estimated camera pose, matches the input image, thereby removing shape mismatch. Finally, unprojecting the input image and the deformed multiview images onto the aligned mesh replaces the blurry colors predicted by the reconstruction model with crisp source pixels. The paper presents this as a new application of deformation fields to the one-image-to-3D task, rather than a new network architecture.

Load-bearing premise

The fidelity module assumes that matching rendered color to the input image is enough to find the right geometric deformation; when differently shaped parts share a color, the optimizer moves the wrong part.

Editorial extensions

If this is right

  • The two deformation modules can be plugged into existing one-image-to-3D methods at inference time and improve their quality, as demonstrated on TripoSR, Unique3D, LGM, and other backbones.
  • Unprojecting the input and deformed multiview images onto the mesh yields sharper colors than letting the reconstruction model predict mesh colors, because the original high-frequency pixels are preserved.
  • The appearance loop removes ghosting caused by multiview inconsistency, making direct image-to-mesh color mapping viable.
  • The fidelity loop makes the mesh geometry match the input view, enabling safe unprojection of the input photo and improving shape fidelity.
  • Total runtime on an NVIDIA A100 with the InstantMesh backbone is about 62 seconds, so the refinement is cheap enough for practical use.

Reading between the lines

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

  • If the unprojection insight generalizes, any multi-view-to-3D method that still predicts texture with a learned decoder could benefit from a similar deferred-coloring step, even without deformation.
  • The color-only supervision of the fidelity module could be extended with semantic or feature-based losses, such as matching segmentations or deep visual features, to avoid the color-ambiguity failures illustrated by the bird-and-branch case.
  • Because the modules are plug-and-play, stacking them on top of future stronger diffusion and reconstruction models should compound quality gains as the baseline improves.
  • A testable prediction is that the 2D appearance module's benefit grows as multiview inconsistency increases; one could verify this by corrupting multiview images with known positional jitter and measuring how FID and LPIPS improve.
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

3 major / 6 minor

Summary. The paper proposes Fancy123, a plug-and-play inference-time enhancement pipeline for single-image-to-3D mesh generation. Starting from an initial mesh obtained via a multiview diffusion model and an LRM (InstantMesh), it adds (1) an appearance enhancement module that deforms the 2D multiview images with grid-based deformation fields to improve multiview consistency, then unprojects them onto the mesh for clear coloration; and (2) a fidelity enhancement module that deforms the 3D mesh via a Jacobian field so that its rendering matches the input image, then unprojects the input image onto the deformed mesh. The authors report state-of-the-art quantitative results on 100 GSO objects, provide qualitative comparisons, ablations, backbone-replacement experiments, and a documented failure case. The central claim is that these two modules, together with an unprojection operation, yield high-quality meshes with better clarity and input fidelity than existing baselines.

Significance. If the claims are substantiated, the paper makes a useful engineering contribution: the two modules are method-agnostic, operate at inference time, and can be dropped into existing one-image-to-3D systems. The qualitative results are often visually compelling, and the paper is commendably transparent, including a detailed failure case in Appendix E and a candid discussion in Sec. 4.1 that the quantitative metrics 'do not align well with human perception.' The release of code and the backbone-replacement experiments (Sec. 4.2, Fig. 7, Appendix C) strengthen reproducibility and demonstrate broad applicability. However, the significance is currently limited by two issues: the headline SoTA claim rests on metrics the authors themselves argue are unreliable, and the fidelity module has an acknowledged identifiability failure that the manuscript does not resolve. With additional validation (e.g., a human study or a constrained objective that avoids the color-ambiguity failure), the work could be an important practical recipe for improving one-image-to-3D pipelines.

major comments (3)
  1. [Sec. 4.1, Appendix B.2, Table 1] The manuscript reports Fancy123 as achieving the best scores on 6 of 7 metrics in Table 1, yet Sec. 4.1 and Appendix B.2 explicitly state that these metrics 'do not align well with human perception' and provide counterexamples (Fig. 5, Fig. 8) where the metric ranking contradicts visual judgment. The paper then recommends relying on qualitative comparisons, but the abstract and introduction still claim 'SoTA performance with significant improvement' based on the same metrics. To make the central claim load-bearing, the authors should either (a) provide a human perceptual evaluation (e.g., pairwise preference with multiple raters and statistical significance) that supports the qualitative superiority, or (b) justify, with error bars and significance tests on the 100-object set, why the existing metrics remain a valid ranking instrument despite the documented mismatches. Without such evidence, the quantitative support for 'significant improvement' is insufficient.
  2. [Sec. 3.5, Appendix E, Fig. 11] The fidelity enhancement module optimizes a Jacobian field using only photometric losses (Eqs. 7-10) from a single viewpoint. The paper's own failure case (Appendix E, Fig. 11) shows that when semantically different parts share similar colors, the optimizer raises the bird's brown back instead of moving the misaligned brown branch, because the colors match. This is not an isolated numerical artifact but an identifiability failure of the photometric objective: many geometrically different meshes render to nearly the same image from one view. The recommendation to skip the 3D-deformation-based fidelity module in such cases effectively concedes that the module is not universally safe. To support the headline claim of reliable high-fidelity improvement, the authors should either add semantic or multi-view constraints to resolve the ambiguity, or explicitly restrict the claim (e.g., by detecting color-ambiguous regions at inference time and disabling the module only when they occur) and demonstrate that the restriction still covers the majority of practical inputs.
  3. [Sec. 3.3, Sec. 4.2, Fig. 2] The pipeline includes a geometry-refinement step (Sec. 3.3, Appendix A.2) that optimizes vertex coordinates against multiview normal maps, yet the abstract and Sec. 1 describe only 'two enhancement modules' and an unprojection operation. The ablation study (Sec. 4.2) isolates the effects of the 2D appearance and 3D fidelity modules, but not the geometry-refinement step, so the contribution of each component to the final result is not fully quantified. Please either include geometry refinement in the contribution list and ablation, or clearly state that it is a preprocessing step borrowed from prior work and, in that case, still measure its impact on the final output.
minor comments (6)
  1. [Sec. 3.3] The phrase 'we adopt an multiview diffusion model' should be 'a multiview diffusion model'.
  2. [Sec. 3.5] The sentence 'so that when rendering Md from π, the result R(Mc, π) would closely resemble Iin' appears to contain a typo: the rendering should be of Md, not Mc, i.e., R(Md, π).
  3. [Eq. (5)] In the definition of Lsmooth2D, the sums over the grid dimensions i and j range from 1 to G-1 and 1 to G respectively, which is clear, but the normalization by K is redundant if K is the number of views; please verify the notation is consistent.
  4. [Table 1] No error bars, confidence intervals, or significance tests are reported for the quantitative comparisons on the 100-object set, which is particularly problematic given the paper's own caveat about metric unreliability.
  5. [Appendix B.1, Table 2] The frontal-view results in Table 2 are reported without statistical significance, and the paper recommends focusing on qualitative results for these too; please consider adding a per-case or paired analysis to support the conclusion.
  6. [Sec. 2, Related Work] The related-work section would benefit from citing the original sources for the GSO dataset evaluation protocol, as well as a clearer discussion of how the proposed unprojection operation differs from the one used in Unique3D.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the headline SoTA claim rests on external GSO benchmarks and human perception, while the two optimization modules explicitly minimize their own stated losses; self-citations are absent.

full rationale

Fancy123 is an inference-time optimization pipeline, and its headline claims are validated externally rather than derived from fitted quantities. The paper contains no self-citations: the Jacobian-field priors [1, 8, 53], the unprojection scheme [49], and the backbone LRM/diffusion models [12, 33, 50] are all external works, and no 'uniqueness theorem' from the authors is invoked. The two enhancement loops are explicit: the appearance loop (Sec. 3.4) optimizes 2D deformation fields so that a mesh colored by the deformed multiview images renders back to those same images (Eqs. 3-6), and the fidelity loop (Sec. 3.5) optimizes a Jacobian field so that the deformed mesh renders to the input image (Eqs. 7-10). One can note a by-construction element: each module's pixel-level success metric equals its own fitted loss, and the final mesh is deformed to and painted from the input image, so at the estimated input view the rendering trivially resembles the input, which is itself a GT render. However, the paper never presents that loss decrease as a prediction; its claims of improved clarity, fidelity, and state-of-the-art quality are supported by independent evidence: GSO-based appearance and geometry metrics (Tab. 1), human perceptual judgments (Figs. 4-8), and plug-and-play application to other backbones (Fig. 7, Tab. 3). The acknowledged failure case (Appendix E, Fig. 11) is an identifiability limit of a photometric-only objective (the optimizer raises a color-matched brown back instead of the misaligned branch), which is a correctness and robustness concern, not circularity. The modest input-view advantage is a small part of the 24-view aggregate and does not make the central claim equivalent to its inputs.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities or forces. Its free parameters are standard optimization hyperparameters (loss weights, grid resolution, iteration counts) chosen by hand. The main axioms are domain assumptions about the adequacy of deformation models and RGBA supervision; one of these assumptions is explicitly contradicted by the paper's own failure case.

free parameters (9)
  • w1 (appearance MSE loss weight) = 1.0
    Empirically set in Appendix A.1; not fitted to data.
  • w2 (appearance mask loss weight) = 1.0
    Empirically set in Appendix A.1.
  • w3 (appearance 2D smooth loss weight) = 0.001
    Empirically set in Appendix A.1.
  • w4 (fidelity MSE loss weight) = 1.0
    Empirically set in Appendix A.1.
  • w5 (fidelity mask loss weight) = 0.1
    Empirically set in Appendix A.1.
  • w6 (fidelity Laplacian smooth loss weight) = 1e5
    Empirically set in Appendix A.1.
  • 2D deformation grid resolution G = 20
    Set by hand in Appendix A.1; controls deformation flexibility.
  • appearance optimization iterations = 100
    Experimentally determined in Sec 3.4.
  • fidelity optimization iterations = 200
    Set by hand in Sec 3.5.
assumptions (4)
  • domain assumption Multiview inconsistency can be corrected by smooth 2D grid deformation fields.
    The appearance module models pixel misalignment as small grid vertex offsets (Sec 3.4); if the inconsistency is large or discontinuous, this deformation model fails.
  • domain assumption RGBA image similarity is a sufficient supervision signal for 3D mesh deformation.
    The fidelity module optimizes only RGB and mask losses against the input image (Sec 3.5); the paper's own failure case (Appendix E) shows this fails when different semantic parts share colors.
  • domain assumption Jacobian-field deformation preserves global cross-view plausibility when supervised from one view.
    The paper argues Jacobian-based deformation is globally smooth (Sec 3.5, Fig 6), but this is only demonstrated qualitatively and is not guaranteed.
  • domain assumption The LRM-initialized mesh is close enough to the true shape that deformation fields can repair appearance without changing topology.
    Both enhancement modules deform an existing mesh; if the initial geometry is severely wrong, no deformation field can fix it and unprojection will still produce artifacts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fancy123: One Image to High-Quality 3D Mesh Generation via Plug-and-Play Deformation." pith.science (2026). https://pith.science/paper/BP45TOHG

@misc{pith2026241116185,
  author       = {Pith},
  title        = {Pith review of: Fancy123: One Image to High-Quality 3D Mesh Generation via Plug-and-Play Deformation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BP45TOHG}},
  note         = {Machine review of arXiv:2411.16185}
}
read the original abstract

Generating 3D meshes from a single image is an important but ill-posed task. Existing methods mainly adopt 2D multiview diffusion models to generate intermediate multiview images, and use the Large Reconstruction Model (LRM) to create the final meshes. However, the multiview images exhibit local inconsistencies, and the meshes often lack fidelity to the input image or look blurry. We propose Fancy123, featuring two enhancement modules and an unprojection operation to address the above three issues, respectively. The appearance enhancement module deforms the 2D multiview images to realign misaligned pixels for better multiview consistency. The fidelity enhancement module deforms the 3D mesh to match the input image. The unprojection of the input image and deformed multiview images onto LRM's generated mesh ensures high clarity, discarding LRM's predicted blurry-looking mesh colors. Extensive qualitative and quantitative experiments verify Fancy123's SoTA performance with significant improvement. Also, the two enhancement modules are plug-and-play and work at inference time, allowing seamless integration into various existing single-image-to-3D methods. Code at: https://github.com/YuQiao0303/Fancy123

Figures

Figures reproduced from arXiv: 2411.16185 by the authors.

Figure 1
Figure 1. Given a single image, our Fancy123 generates a high-quality 3D mesh in terms of visual appeal, color clarity, and input fidelity. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Our pipeline: (1) Initialization via multiview diffusion and LRM: given an input image I in, use a multiview (MV) diffusion model to generate MV images {Ik}, use an LRM to create a mesh conditioned on {Ik}, and then refine mesh geometry to obtain the initial mesh M0. (2) Appearance enhancement: optimize 2D deformation fields to deform {Ik} to {I ′ k} to ensure MV consistency when unprojecting {I ′ k} onto M0 to get … view at source ↗
Figure 3
Figure 3. 2D Deformation Optimization Loop: optimize 2D de [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Visual comparisons of our method against baseline methods for the single-image-to-3D-mesh task. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Ablation experiments on our appearance (a-d) and fidelity (e-i) enhancement modules via 2D and 3D deformation, respectively. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Superiority of Jacobian-based deformation: globally [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Improvement is significant when applying our various [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Illustration of the first limitation of existing quantitative metrics: mismatch with human perception. From (a) to (h), each sample [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Illustration of the second limitation of existing quantitative metrics: the contradiction between the ambiguous task and the unique [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 11
Figure 11. Figure 11: A failure case of our Fancy123 when different semantic [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: More visual comparisons. Our method exhibits high clarity and plausibility compared to baseline methods. Better zoom in. [PITH_FULL_IMAGE:figures/full_fig_p019_12.png]
Figure 13
Figure 13. Figure 13: More visual comparisons. Our method exhibits high clarity and plausibility compared to baseline methods. Better zoom in. [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 14
Figure 14. Figure 14: Ablation experiments on the 2D-deformation-based appearance enhancement module: unprojecting the deformed multiview [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]
Figure 15
Figure 15. Figure 15: Ablation experiments on the 2D-deformation-based appearance enhancement module: unprojecting the deformed multiview [PITH_FULL_IMAGE:figures/full_fig_p022_15.png]
Figure 16
Figure 16. Figure 16: Ablation experiments on 3D mesh deformation. For simplicity, [PITH_FULL_IMAGE:figures/full_fig_p023_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 39 canonical work pages

  1. [1]

    Neural Jaco- bian Fields: learning intrinsic mappings of arbitrary meshes

    Noam Aigerman, Kunal Gupta, Vladimir G Kim, Siddhartha Chaudhuri, Jun Saito, and Thibault Groueix. Neural Jaco- bian Fields: learning intrinsic mappings of arbitrary meshes. ACM Transactions on Graphics (TOG), 41(4):1–17, 2022. 2, 3, 4

    work page 2022

  2. [2]

    Unique3d/app/custom models/normal prediction.py at main

    AiuniAI. Unique3d/app/custom models/normal prediction.py at main. https : / / github . com / AiuniAI / Unique3D / blob / main / app / custom _ models / normal_prediction.py, 2024. 1

    work page 2024

  3. [3]

    SF3D: Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement

    Mark Boss, Zixuan Huang, Aaryaman Vasishta, and Varun Jampani. SF3D: Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement. arXiv preprint arXiv:2408.00653, 2024. 6, 2

    arXiv 2024

  4. [4]

    Real-time deformable reg- istration for robotic ultrasound surgical guidance based on vectorized trust-region optimization

    Teru Chen, Xingwei Zhao, Guo Zheng, Jiaxin Wang, Qing Ling, Bo Tao, and Zhouping Yin. Real-time deformable reg- istration for robotic ultrasound surgical guidance based on vectorized trust-region optimization. IEEE Transactions on Instrumentation and Measurement, 73:1–11, 2024. 2, 4

    work page 2024

  5. [5]

    Learning implicit fields for generative shape modeling

    Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, pages 5939–5948,

  6. [6]

    Objaverse-xl: A universe of 10m+ 3d objects

    Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Chris- tian Laforte, Vikram V oleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl V ondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objaverse-xl: A universe of 10m+ 3d objects. arXiv preprint arXiv:2307.05663, 2023. 2

    arXiv 2023

  7. [7]

    Google scanned objects: A high- quality dataset of 3d scanned household items

    Laura Downs, Anthony Francis, Nate Koenig, Brandon Kin- man, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high- quality dataset of 3d scanned household items. In Inter- national Conference on Robotics and Automation (ICRA) , pages 2553–2560. IEEE, 2022. 6

    work page 2022

  8. [8]

    TextDeformer: Geometry Manipu- lation using Text Guidance

    William Gao, Noam Aigerman, Groueix Thibault, Vladimir Kim, and Rana Hanocka. TextDeformer: Geometry Manipu- lation using Text Guidance. InACM Transactions on Graph- ics (SIGGRAPH), 2023. 2, 3, 4

    work page 2023

Show all 57 references
  1. [9]

    Shape and viewpoints without keypoints

    Shubham Goel, Angjoo Kanazawa, and Jitendra Malik. Shape and viewpoints without keypoints. In ECCV, 2020. 3

    2020

  2. [10]

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

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equi- librium. Advances in neural information processing systems, 30, 2017. 6

    2017

  3. [11]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 2

    2020

  4. [12]

    LRM: Large reconstruction model for single image to 3D

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In ICLR, 2024. 1, 2

    2024

  5. [13]

    Self-supervised 3d mesh reconstruction from single images

    Tao Hu, Liwei Wang, Xiaogang Xu, Shu Liu, and Jiaya Jia. Self-supervised 3d mesh reconstruction from single images. In CVPR, pages 6002–6011, 2021. 3

    2021

  6. [14]

    As- rigid-as-possible shape manipulation

    Takeo Igarashi, Tomer Moscovich, and John F Hughes. As- rigid-as-possible shape manipulation. ACM transactions on Graphics (TOG), 24(3):1134–1141, 2005. 3

    2005

  7. [15]

    Keypointdeformer: Unsupervised 3D keypoint discovery for shape control

    Tomas Jakab, Richard Tucker, Ameesh Makadia, Jiajun Wu, Noah Snavely, and Angjoo Kanazawa. Keypointdeformer: Unsupervised 3D keypoint discovery for shape control. In CVPR, pages 12783–12792, 2021. 3

    2021

  8. [16]

    Efros, and Jitendra Malik

    Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruc- tion from image collections. In ECCV, 2018. 3

    2018

  9. [17]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4), 2023. 3

    2023

  10. [18]

    OptCtrlPoints: Finding the Optimal Control Points for Bi- harmonic 3D Shape Deformation

    Kunho Kim, Mikaela Angelina Uy, Despoina Paschalidou, Alec Jacobson, Leonidas J Guibas, and Minhyuk Sung. OptCtrlPoints: Finding the Optimal Control Points for Bi- harmonic 3D Shape Deformation. In Computer Graphics Forum, page e14963. Wiley Online Library, 2023. 3

    2023

  11. [19]

    LN3Diff: Scalable Latent Neural Fields Diffusion for Speedy 3D Generation

    Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. LN3Diff: Scalable Latent Neural Fields Diffusion for Speedy 3D Generation. In ECCV, 2024. 2

    2024

  12. [20]

    Self-supervised Single-view 3D Reconstruction via Semantic Consistency

    Xueting Li, Sifei Liu, Kihwan Kim, Shalini De Mello, Varun Jampani, Ming-Hsuan Yang, and Jan Kautz. Self-supervised Single-view 3D Reconstruction via Semantic Consistency. In ECCV, 2020. 3

    2020

  13. [21]

    Magic3d: High-resolution text-to-3d content creation

    Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 5

    2023

  14. [22]

    Linear rotation-invariant coordinates for meshes

    Yaron Lipman, Olga Sorkine, David Levin, and Daniel Cohen-Or. Linear rotation-invariant coordinates for meshes. ACM Transactions on Graphics (ToG) , 24(3):479–487,

  15. [23]

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

    Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tok- makov, Sergey Zakharov, and Carl V ondrick. Zero-1-to-3: Zero-shot one image to 3D object. In ICCV, pages 9298– 9309, 2023. 2

    2023

  16. [24]

    SyncDreamer: Gen- erating Multiview-consistent Images from a Single-view Im- age

    Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. SyncDreamer: Gen- erating Multiview-consistent Images from a Single-view Im- age. In ICLR, 2024. 1, 2, 3, 6

    2024

  17. [25]

    Wonder3D: Sin- gle Image to 3D using Cross-Domain Diffusion

    Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3D: Sin- gle Image to 3D using Cross-Domain Diffusion. In CVPR,

  18. [26]

    Mattes, D.R

    D. Mattes, D.R. Haynor, H. Vesselle, T.K. Lewellen, and W. Eubank. PET-CT image registration in the chest using free- form deformations. IEEE Transactions on Medical Imaging, 22(1):120–128, 2003. 2, 4

    2003

  19. [27]

    Occupancy Net- works: Learning 3D reconstruction in function space

    Lars Mescheder, Michael Oechsle, Michael Niemeyer, Se- bastian Nowozin, and Andreas Geiger. Occupancy Net- works: Learning 3D reconstruction in function space. In CVPR, pages 4460–4470, 2019. 3

    2019

  20. [28]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 3

    2020

  21. [29]

    Yatian Pang, Tanghui Jia, Yujun Shi, Zhenyu Tang, Junwu Zhang, Xinhua Cheng, Xing Zhou, Francis E. H. Tay, and Li Yuan. Envision3d: One image to 3d with anchor views interpolation, 2024. 2

    2024

  22. [30]

    DeepSDF: Learning Continuous Signed Distance Functions for Shape Represen- tation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning Continuous Signed Distance Functions for Shape Represen- tation. In CVPR, 2019. 3

    2019

  23. [31]

    Barron, and Ben Milden- hall

    Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Milden- hall. DreamFusion: Text-to-3D using 2D Diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 5

    2022 arXiv

  24. [32]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  25. [33]

    Zero123++: a single image to consistent multi-view dif- fusion base model

    Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view dif- fusion base model. arXiv preprint arXiv:2310.15110, 2023. 2, 3, 4, 1

    arXiv

  26. [34]

    MVDream: Multi-view Diffusion for 3D Generation

    Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. MVDream: Multi-view Diffusion for 3D Generation. arXiv:2308.16512, 2023. 1, 2, 3

    2023 arXiv

  27. [35]

    Laplacian surface editing

    Olga Sorkine, Daniel Cohen-Or, Yaron Lipman, Marc Alexa, Christian R ¨ossl, and H-P Seidel. Laplacian surface editing. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH symposium on Geometry processing , pages 175–184, 2004. 3

    2004

  28. [36]

    zero123plus/examples/normal gen.py at main

    SUDO-AI-3D. zero123plus/examples/normal gen.py at main. https : / / github . com / SUDO - AI - 3D / zero123plus / blob / main / examples / normal _ gen.py, 2024. 1

    2024

  29. [37]

    Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior

    Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In ICCV, pages 22819–22829, 2023. 5

    2023

  30. [38]

    LGM: Large Multi-View Gaus- sian Model for High-Resolution 3D Content Creation

    Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large Multi-View Gaus- sian Model for High-Resolution 3D Content Creation. In ECCV, pages 1–18. Springer, 2024. 1, 6, 7, 8, 2

    2024

  31. [39]

    Dreamgaussian: Generative gaussian splatting for ef- ficient 3d content creation

    Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for ef- ficient 3d content creation. In ICLR, 2024. 5

    2024

  32. [40]

    Cy- cle3D: High-quality and Consistent Image-to-3D Genera- tion via Generation-Reconstruction Cycle

    Zhenyu Tang, Junwu Zhang, Xinhua Cheng, Wangbo Yu, Chaoran Feng, Yatian Pang, Bin Lin, and Li Yuan. Cy- cle3D: High-quality and Consistent Image-to-3D Genera- tion via Generation-Reconstruction Cycle. arXiv preprint arXiv:2407.19548, 2024. 2, 3

    2024 arXiv

  33. [41]

    Digital twin modeling

    Fei Tao, Bin Xiao, Qinglin Qi, Jiangfeng Cheng, and Ping Ji. Digital twin modeling. Journal of Manufacturing Systems , 64:372–389, 2022. 1

    2022

  34. [42]

    TripoSR: Fast 3D Object Reconstruction from a Single Image

    Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. TripoSR: Fast 3D Object Reconstruction from a Single Image. arXiv preprint arXiv:2403.02151, 2024. 6, 8, 2

    2024 arXiv

  35. [43]

    Pixel2Mesh: Generating 3D Mesh Models from Single RGB Images

    Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2Mesh: Generating 3D Mesh Models from Single RGB Images. In ECCV, 2018. 3

    2018

  36. [44]

    ImageDream: Image-Prompt Multi-view Diffusion for 3D Generation

    Peng Wang and Yichun Shi. ImageDream: Image-Prompt Multi-view Diffusion for 3D Generation. arXiv preprint arXiv:2312.02201, 2023. 1, 2

    2023 arXiv

  37. [45]

    NeuS: Learning Neural Im- plicit Surfaces by V olume Rendering for Multi-view Recon- struction

    Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Im- plicit Surfaces by V olume Rendering for Multi-view Recon- struction. Advances in Neural Information Processing Sys- tems, 34:27171–27183, 2021. 1, 2

    2021

  38. [46]

    A survey on metaverse: Fundamentals, security, and privacy

    Yuntao Wang, Zhou Su, Ning Zhang, Rui Xing, Dongxiao Liu, Tom H Luan, and Xuemin Shen. A survey on metaverse: Fundamentals, security, and privacy. IEEE Communications Surveys & Tutorials, 25(1):319–352, 2022. 1

    2022

  39. [47]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600–612, 2004. 6

    2004

  40. [48]

    CRM: Single image to 3D textured mesh with con- volutional reconstruction model

    Zhengyi Wang, Yikai Wang, Yifei Chen, Chendong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. CRM: Single image to 3D textured mesh with con- volutional reconstruction model. In ECCV, pages 57–74. Springer, 2025. 1, 6, 2

    2025

  41. [49]

    Unique3D: High-Quality and Efficient 3D Mesh Generation from a Single Image

    Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3D: High-Quality and Efficient 3D Mesh Generation from a Single Image. In NeurIPS, 2024. 2, 3, 4, 5, 6, 7, 8, 1

    2024

  42. [50]

    InstantMesh: Effi- cient 3D Mesh Generation from a Single Image with Sparse-view Large Reconstruction Models

    Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. InstantMesh: Effi- cient 3D Mesh Generation from a Single Image with Sparse-view Large Reconstruction Models. arXiv preprint arXiv:2404.07191, 2024. 4, 5, 6, 7, 1, 2

    2024 arXiv

  43. [51]

    Hi3D: Pursuing High- Resolution Image-to-3D Generation with Video Diffusion Models

    Haibo Yang, Yang Chen, Yingwei Pan, Ting Yao, Zhineng Chen, Chong-Wah Ngo, and Tao Mei. Hi3D: Pursuing High- Resolution Image-to-3D Generation with Video Diffusion Models. In ACM MM, 2024. 2

    2024

  44. [52]

    ConsistNet: Enforcing 3D Consistency for Multi- view Images Diffusion

    Jiayu Yang, Ziang Cheng, Yunfei Duan, Pan Ji, and Hong- dong Li. ConsistNet: Enforcing 3D Consistency for Multi- view Images Diffusion. arXiv preprint arXiv:2310.10343 ,

    arXiv

  45. [53]

    Kim, and Minhyuk Sung

    Seungwoo Yoo, Kunho Kim, Vladimir G. Kim, and Minhyuk Sung. As-Plausible-As-Possible: Plausibility-Aware Mesh Deformation Using 2D Diffusion Priors. In CVPR, 2024. 2, 3, 4

    2024

  46. [54]

    Repaint123: Fast and high-quality one image to 3d genera- tion with progressive controllable 2d repainting

    Junwu Zhang, Zhenyu Tang, Yatian Pang, Xinhua Cheng, Peng Jin, Yida Wei, Wangbo Yu, Munan Ning, and Li Yuan. Repaint123: Fast and high-quality one image to 3d genera- tion with progressive controllable 2d repainting. In ECCV,

  47. [55]

    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, pages 586–595,

  48. [57]

    Camera pose estimation

    We empirically set our loss weights w1-w6 for LMSE1, Lmask1, Lsmooth2D, LMSE2, Lmask2, LLap to be 1.0, 1.0, 0.001, 1.0, 0.1, and 1e5, respectively. During the quantitative eval- uation, the threshold for F-score is set to 0.2 following In- stantMesh [50]. A.2. Geometry refinem...

  49. [2018]

    Implementation details A.1

    5, 6 Fancy123: One Image to High-Quality 3D Mesh Generation via Plug-and-Play Deformation Supplementary Material A. Implementation details A.1. Hyper parameters We use the same camera settings as InstantMesh [50]. The default camera FOV (Field of View) angle is 30°, and the de...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.