Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Zero-P-to-3: Zero-Shot Partial-View Images to 3D Object

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Zero-P-to-3 reconstructs complete 3D objects from dense views covering only a 90-degree arc by fusing three priors inside DDIM sampling.

desk verdict A training-free partial-view 3D reconstruction worth refereeing, but the printed evaluation oversells the gains while the key schedule equation contradicts its own prose. read the letter →

arxiv 2505.23054 v1 pith:6KUXHDF5 submitted 2025-05-29 cs.CV

Yuxuan Lin , Ruihang Chu , Zhenyu Chen , Xiao Tang , Lei Ke , Haoling Li , Yingji Zhong , Zhihao Li
show 4 more authors
Shiyong Liu Xiaofei Wu Jianzhuang Liu Yujiu Yang
This is my paper
classification cs.CV
keywords partial-view3Dreconstructionzero-shotGaussianSplattingmulti-viewdiffusionDDIMsamplingimagerestorationpriornovelviewsynthesisconsistency
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 claims that a complete 3D object can be reconstructed from dense images covering only about 90 degrees of viewing angle, without any training of a new model. Its proposed pipeline first builds a coarse 3D Gaussian Splatting model from the visible views, then uses a fusion-based DDIM sampling procedure to generate consistent images for the unseen sides, and finally refines the model with rotated-view supervision. The central argument is that this succeeds because the sampling combines three complementary signals—a multi-view diffusion prior, a low-frequency geometric prior taken from the coarse model's rendering, and a high-frequency image-restoration prior—with time-dependent weights that shift from structure to detail during denoising. If correct, this means existing pretrained generative components are enough to repair invisible regions coherently while preserving the fidelity of the observed regions, which matters for everyday captures of objects blocked by furniture, walls, or other occluders.

What carries the argument

The load-bearing mechanism is the fused noise prediction used at every DDIM step, formally $\varepsilon^t = \varepsilon_{\mathrm{MVD}} + w_{\mathrm{HF}}(t)\varepsilon^t_{\mathrm{HF}} + w_{\mathrm{LF}}(t)\varepsilon^t_{\mathrm{LF}}$, where $\varepsilon_{\mathrm{MVD}}$ is the weighted average of multi-view diffusion predictions, $\varepsilon^t_{\mathrm{LF}}$ is the noise obtained by diffusing the coarse 3DGS render at the target view, and $\varepsilon^t_{\mathrm{HF}}$ comes from an image-restoration diffusion model. The time-dependent weights, $w_{\mathrm{LF}}(t)=\frac{\tanh(-(t-\tau)/\sigma)+1}{2}$ and $w_{\mathrm{HF}}(t)=\frac{1-w_{\mathrm{LF}}(t)}{\eta}$, make geometry dominate early in denoising and texture refinement dominate late. The rotated-view refinement then densifies supervision by generating images at incrementally rotated angles and optimizing the 3DGS parameters with an L1 plus LPIPS loss.

What would settle it

Render a known object from a 90-degree arc, build the coarse model, then replace its invisible-region renderings with a deliberately wrong shape before running the refinement. If the final invisible-region reconstruction follows the wrong shape, the geometric prior is the active mechanism and its unseen-region reliability is the assumption to check; if it ignores the wrong shape, the geometric prior is not what is carrying the gain. Separately, measuring the coarse model's invisible-region error before refinement and correlating it with the reported 18.14 PSNR would show whether the fusion is correcting the coarse geometry or merely painting over it.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the failure mode of multi-view diffusion for partial observations—blurriness and conflicting content in unobserved regions—stems from treating reference views as conditionally independent when averaging their score predictions. Zero-P-to-3 replaces the plain averaged score with a fused noise prediction inside DDIM sampling, adding a coarse 3DGS rendering as a low-frequency geometric anchor and an image-restoration prior as a high-frequency texture cue. It then rotates the target viewpoints in batches around the observed angular range, using the fused sampling to produce supervision images and optimizing the Gaussian model with reconstruction and perceptual losses. In the paper's experiments on a synthetic object benchmark with ground-truth views, this yields the strongest overall and invisible-region metrics among the compared reconstruction and generative baselines, with invisible-region PSNR of 18.14 against 17.63 for plain 3DGS.

Load-bearing premise

The load-bearing premise is that the coarse 3D model built from the visible views is a trustworthy guide to the geometry of the parts no camera ever saw, because the fused sampling is pulled toward that model's renderings more and more strongly in the later denoising steps.

Editorial extensions

If this is right

  • Objects observed only from a narrow angular span get plausible, view-consistent back and side geometry without training a new network.
  • The same fused-sampling recipe can be attached to any off-the-shelf multi-view diffusion model and any coarse 3D representation, since the geometric and restoration priors enter only at inference time.
  • Rotated-view supervision reduces the anisotropy artifacts of 3DGS when the same scene is viewed from novel angles.
  • Because visible-region fidelity remains close to plain 3DGS while invisible regions improve, the method splits the difference between reconstruction-based and generative-based pipelines.

Reading between the lines

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

  • An implicit extension is that the frequency-decomposed fusion recipe could be reused with other renderer/diffusion pairs, such as mesh rasterizers or other point-based renderers, by simply replacing the coarse renderer and the restoration network.
  • A stress test for the mechanism would be to deliberately corrupt the coarse model's invisible-region renderings; if the final reconstruction tracks the corruption, the geometric prior is the dominant source of the invisible-region gains, and if it resists, the diffusion prior is doing the heavy lifting.
  • The method's success on a 90-degree arc raises a scaling question beyond the paper: how the quality of hidden regions degrades as the observed arc shrinks below 90 degrees or becomes fragmented by multiple occluders.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Zero-P-to-3, a training-free pipeline for 3D reconstruction from dense partial views covering only a limited angular range (about 90 degrees). The method first fits a coarse 3D Gaussian Splatting (3DGS) model to the observed views, then generates novel-view images by DDIM sampling in which three signals are fused: a multi-view diffusion prediction (Era3D), a low-frequency geometric rendering prior computed from the coarse 3DGS model, and a high-frequency image-restoration prior (DiffBIR). The resulting images are used as supervision in an iterative refinement loop that rotates novel viewpoints around the object and optimizes 3DGS parameters with L1 and LPIPS losses. Experiments on Objaverse and real-world data report state-of-the-art results, including a total PSNR of 18.74 dB and an invisible-region PSNR of 18.14 dB, the latter compared with 17.63 dB for plain 3DGS.

Significance. If the claims hold, the paper makes a useful contribution to an underexplored setting: dense but angularly restricted observations, where traditional sparse-view interpolation fails and single-view generative methods discard the available multi-view information. The modular design—combining a multi-view diffusion prior, a geometric anchor, and an image-restoration prior inside DDIM—is a sensible and potentially reusable idea, and the paper should be credited for providing a clear ablative breakdown, including a monotonic progression in Table 2, and for evaluating on both synthetic and real-world data. The reported gains are internally coherent, and the qualitative figures support the general direction. However, several load-bearing aspects of the evaluation and of the fusion schedule need to be fixed before the central claims can be accepted.

major comments (4)
  1. [Sec. 3.3.2, Eq. (6), and Sec. 4.1] The stated transition schedule contradicts the accompanying prose. With the printed hyperparameters tau=22, sigma=7 and DDIM steps t=50,...,1, Eq. (6) gives w_LF(50) ~= 0.0003, w_LF(22)=0.5, and w_LF(1) ~= 0.9975, meaning the low-frequency geometric prior is imposed hardest in the final denoising steps, not at large t. The text in Sec. 3.3.2 says "geometric structure dominates at large t while texture refinement becomes prominent at small t," which is exactly the opposite behavior. This matters because the final steps determine fine structure: if the coarse model's invisible-region geometry is unreliable, the schedule forces the generated images toward exactly the errors the refinement is supposed to repair. Please correct the schedule, the prose, or the indexing convention, and show the actual weight curves used in the experiments.
  2. [Table 2, rows 6 and 8, and Table 1] The ablation study does not isolate the effect of the geometric prior on invisible regions. Comparing row 6 (without the geometric prior) and row 8 (Ours), the total PSNR gain is only 0.08 dB, and no visible/invisible breakdown is given for any ablation row. The main result in Table 1 attributes the invisible-region improvement over 3DGS (18.14 vs. 17.63 dB) to the full pipeline, but on the evidence of Table 2 the geometric prior contributes very little overall. Without an invisible-region ablation, the reader cannot tell whether the claimed gain comes from the geometric anchor correcting the coarse model, from the diffusion prior hallucinating plausible texture, or from the rotated-view refinement. Please report visible/invisible PSNR, SSIM, and LPIPS for each ablation configuration.
  3. [Sec. 3.3.2, Eq. (4)] The geometric prior x_LF is rendered from the coarse 3DGS model G0, which is fitted only to the visible 90-degree arc. In invisible regions, x_LF is not measured structure but an unconstrained product of the 3DGS optimization. The paper provides no quantitative evaluation of G0 alone at unseen viewpoints, so there is no evidence that this placeholder is a trustworthy low-frequency anchor. Given the schedule issue noted above, the method can propagate errors from G0 into the final generated views. Please add an evaluation of the coarse model's novel-view quality and, ideally, a sensitivity analysis showing how final reconstruction quality varies with the quality of the coarse anchor.
  4. [Sec. 4.2, experimental setup] The main comparison is not conducted under an equal input budget. Zero-P-to-3 and 3DGS use 90 input images, GaussianObject and EscherNet use 30, Trellis uses 5, and Era3D/LGM/OpenLRM use 1. The headline superiority over generative single-view baselines may therefore be largely attributable to the much larger amount of observed data. Please either rerun the comparison with matched input counts for all methods or explicitly quantify how performance changes as the number of input views to Zero-P-to-3 is reduced.
minor comments (5)
  1. [Sec. 2] The section heading "Reated Work" contains a typo; it should read "Related Work."
  2. [Sec. 3.3, Eq. (2)] The displayed conditional-independence expression is malformed: it writes p(x_t|{V_i}) ∝ Q_i p(x_t|{V_i}^N_i=1), which mixes the whole set and an element. It should be p(x_t|{V_i}) ∝ ∏_i p(x_t|V_i), assuming conditional independence.
  3. [Sec. 3.3] The convention for the diffusion time index is confusing: the text says denoising progresses "from pure noise (t=1) to a clean image (t=0)," while Eq. (7) and the 50-step DDIM schedule seem to use a different indexing. Please define the time direction explicitly and use it consistently in Eq. (6) and Eq. (7).
  4. [Sec. 4.1 and Table 1] The description of the real-world dataset is very brief, and Figure 6 shows "no GT" images without specifying the evaluation protocol used for those cases (visual inspection only or some perceptual metric). Please state exactly how real-world results were assessed.
  5. [Sec. 3.3.1] The text states the weights are set to 2, 1.5, and 1 for frontal, back, and other views, but the method description does not explain how "frontal" and "back" are identified relative to the partial-view input arc. Please define this more precisely.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the coarse-3DGS geometric prior is a regularizer, not a prediction, and the reported gains are measured against external ground-truth images.

full rationale

The only candidate for a circular step is the geometric prior in Sec. 3.3.2 (Eq. 4), where x_LF is a rendering of the coarse 3DGS model G0 fit to the input views, and the iterative refinement in Sec. 3.4 uses diffusion-generated images, conditioned on x_LF, as supervision to refine G0. For this to be circular, the generated images would have to reduce to x_LF by construction. They do not: Eq. 5 combines the multi-view diffusion prediction ε_MVD, the restoration prior ε_HF, and ε_LF with time-dependent weights, and the full method's total PSNR (18.74) differs from plain 3DGS (18.38) and from the GP-only ablation (18.38), while the no-GP variant is 18.66, so the output is not forced to equal the coarse fit. External content enters through the pretrained Era3D multi-view diffusion and DiffBIR restoration priors, which are not fitted to this paper's input views. There is no load-bearing self-citation; the only co-authored reference (CVT-xRF [42]) is related-work context. The paper explicitly acknowledges G0 'typically exhibits incomplete details and inconsistencies in occluded or unseen regions' (Sec. 3.2), so relying on it as a prior is a stated limitation. The w_LF schedule in Eq. 6 (τ=22, σ=7) makes w_LF near zero at large t and near one at small t, contradicting the text's claim that geometry dominates at large t; this is a correctness/robustness concern (the anchor is applied hardest in the final denoising steps), but it does not make the evaluation reduce to the fit, so it is not circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method is a composition of pretrained and hand-set ingredients: the only external information about invisible regions comes from the Era3D and DiffBIR priors, while the geometric prior re-injects the fit of the current 3DGS model into the sampling process. The hand-chosen schedule parameters, view weights, rotation angles, and loss weight are not sensitivity-analyzed, and the printed schedule contradicts the prose describing it. No new physical or conceptual entities are postulated; the geometric and restoration priors are existing models treated as modules.

free parameters (5)
  • LF/HF fusion schedule parameters tau, sigma, eta = tau=22, sigma=7, eta=4
    Hand-set transition point, smoothness, and high-frequency scaling in Eq. 6; no sensitivity analysis, and the printed schedule behaves opposite to its prose.
  • Multi-view fusion view weights w_i = frontal 2, back 1.5, others 1
    Hand-assigned weights in Eq. 3; ambiguous because the partial-observation input set contains no back views.
  • Rotation angle Delta_theta and step delta_e = Delta_theta=45, delta_e=6
    Hand-set angular offsets for rotated supervision views in Fig. 5; no ablation on these values.
  • Loss weight lambda = 1
    Balances L_rec and LPIPS in Eq. 8; no sensitivity analysis.
  • Refinement protocol = 3000 steps per iteration, densification in first 1300 steps
    Number of refinement iterations is never stated; the step counts are procedural choices without justification.
assumptions (4)
  • domain assumption The coarse 3DGS model optimized from partial views is a trustworthy geometric anchor for unseen regions.
    The low-frequency prior (Eq. 4) injects a rendering of regions that were never observed into the sampling; if the coarse geometry is wrong there, the generated supervision inherits the error. Entered in Sec. 3.3.2.
  • domain assumption Averaging per-view diffusion scores approximates the true conditional score once regularized.
    Eqs. 2 and 3 assume conditional independence of reference views; the paper states this is invalid and only compensates with heuristic priors, and no corrected approximation is derived. Sec. 3.3.
  • domain assumption Pretrained Era3D and DiffBIR priors transfer to partial-observation views rendered from orthographic cameras at 100x distance.
    These priors are the only information source for invisible-region content; the extreme camera adjustment in Sec. 4.1 is not validated separately.
  • standard math Standard DDIM sampling and COLMAP pose estimation behave as expected.
    Background machinery used without verification in this setting; reasonable but not addressed by the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Zero-P-to-3: Zero-Shot Partial-View Images to 3D Object." pith.science (2026). https://pith.science/paper/6KUXHDF5

@misc{pith2026250523054,
  author       = {Pith},
  title        = {Pith review of: Zero-P-to-3: Zero-Shot Partial-View Images to 3D Object},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6KUXHDF5}},
  note         = {Machine review of arXiv:2505.23054}
}
read the original abstract

Generative 3D reconstruction shows strong potential in incomplete observations. While sparse-view and single-image reconstruction are well-researched, partial observation remains underexplored. In this context, dense views are accessible only from a specific angular range, with other perspectives remaining inaccessible. This task presents two main challenges: (i) limited View Range: observations confined to a narrow angular scope prevent effective traditional interpolation techniques that require evenly distributed perspectives. (ii) inconsistent Generation: views created for invisible regions often lack coherence with both visible regions and each other, compromising reconstruction consistency. To address these challenges, we propose \method, a novel training-free approach that integrates the local dense observations and multi-source priors for reconstruction. Our method introduces a fusion-based strategy to effectively align these priors in DDIM sampling, thereby generating multi-view consistent images to supervise invisible views. We further design an iterative refinement strategy, which uses the geometric structures of the object to enhance reconstruction quality. Extensive experiments on multiple datasets show the superiority of our method over SOTAs, especially in invisible regions.

Figures

Figures reproduced from arXiv: 2505.23054 by the authors.

Figure 1
Figure 1. Zero-P-to-3 reconstructs 3D objects where visible regions are confined to a limited field-of-view. We first reconstruct a coarse 3D [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Visualizations of input images (top), masks (middle), [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of Zero-P-to-3 for 3D reconstruction from partial observations. Starting with input partial images, the system first [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Schematic diagram of rotated views. Yellow lines in [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparisons between our method and other approaches. Reconstruction-based methods (3DGS, GaussianObject) [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: Effects of applying the Image Restoration Prior (IRP) [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 10
Figure 10. Figure 10: Effects of rotated view sampling. The regions framed [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 7
Figure 7. Figure 7: Ablation study on the effect of Multi-view Fusion (MF). [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 11
Figure 11. Figure 11: Analysis of direct reconstruction using results gen [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]
Figure 9
Figure 9. Figure 9: Effects of Geometric Prior in multi-view fusion. Each [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 21 canonical work pages

  1. [1]

    pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction, 2024

    David Charatan, Sizhe Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction, 2024. 1, 2

    work page 2024

  2. [2]

    Objaverse-xl: A universe of 10m+ 3d objects.arXiv preprint arXiv:2307.05663, 2023

    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. 6

    arXiv 2023

  3. [3]

    Depth-supervised nerf: Fewer views and faster train- ing for free

    Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ra- manan. Depth-supervised nerf: Fewer views and faster train- ing for free. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882– 12891, 2022. 2

    2022

  4. [4]

    Dimvis: Diffusion-based multi-view synthesis

    Giuseppe Di Giacomo, Giulio Franzese, Tania Cerquitelli, Carla Fabiana Chiasserini, and Pietro Michiardi. Dimvis: Diffusion-based multi-view synthesis. InICML 2024 Work- shop on Structured Probabilistic Inference{\&}Generative Modeling, 2024. 3

    work page 2024

  5. [5]

    Binocular-guided 3d gaussian splatting with view con- sistency for sparse view synthesis, 2024

    Liang Han, Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Binocular-guided 3d gaussian splatting with view con- sistency for sparse view synthesis, 2024. 2

    work page 2024

  6. [6]

    Viewdiff: 3d-consistent image generation with text-to-image models

    Lukas H ¨ollein, Aljaˇz Boˇziˇc, Norman M¨uller, David Novotny, Hung-Yu Tseng, Christian Richardt, Michael Zollh ¨ofer, and Matthias Nießner. Viewdiff: 3d-consistent image generation with text-to-image models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5043–5052, 2024. 3

    work page 2024

  7. [7]

    Lrm: Large reconstruction model for single image to 3d.arXiv preprint arXiv:2311.04400, 2023

    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.arXiv preprint arXiv:2311.04400, 2023. 2, 6

    arXiv 2023

  8. [8]

    Leap: Liberate sparse-view 3d modeling from camera poses,

    Hanwen Jiang, Zhenyu Jiang, Yue Zhao, and Qixing Huang. Leap: Liberate sparse-view 3d modeling from camera poses,

    work page

Show all 43 references
  1. [9]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

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

  2. [10]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 4015–4026, 2023. 6

    2023

  3. [11]

    Xin Kong, Shikun Liu, Xiaoyang Lyu, Marwan Taher, Xiao- juan Qi, and Andrew J. Davison. Eschernet: A generative model for scalable view synthesis, 2024. 2, 3, 6

    2024

  4. [12]

    Era3d: High-resolution multiview diffusion using efficient row-wise attention.arXiv preprint arXiv:2405.11616, 2024

    Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3d: High-resolution multiview diffusion using efficient row-wise attention.arXiv preprint arXiv:2405.11616, 2024. 1, 2, 3, 6

    2024 arXiv

  5. [13]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2023

  6. [14]

    Re- conx: Reconstruct any scene from sparse views with video diffusion model.arXiv preprint arXiv:2408.16767, 2024

    Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Re- conx: Reconstruct any scene from sparse views with video diffusion model.arXiv preprint arXiv:2408.16767, 2024. 3

    2024 arXiv

  7. [15]

    One-2-3-45++: Fast single im- age to 3d objects with consistent multi-view generation and 3d diffusion

    Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Ji- ayuan Gu, and Hao Su. One-2-3-45++: Fast single im- age to 3d objects with consistent multi-view generation and 3d diffusion. InProceedings of the IEEE/CVF Conference on C...

    2024

  8. [16]

    Syncdreamer: Gen- erating multiview-consistent images from a single-view im- age.arXiv preprint arXiv:2309.03453, 2023

    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.arXiv preprint arXiv:2309.03453, 2023. 1, 3

    2023 arXiv

  9. [17]

    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. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pat...

    2024

  10. [18]

    Reg- nerf: Regularizing neural radiance fields for view synthesis from sparse inputs

    Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg- nerf: Regularizing neural radiance fields for view synthesis from sparse inputs. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, ...

    2022

  11. [19]

    Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022. 3

    2022 arXiv

  12. [20]

    Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors.arXiv preprint arXiv:2306.17843,

    Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Sko- rokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors.arXiv preprint arXiv:2306.17843,

    arXiv

  13. [21]

    High-resolution image syn- thesis with latent diffusion models, 2022

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

    2022

  14. [22]

    Structure- from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 4104–4113, 2016. 3 9

    2016

  15. [23]

    Zero123++: a single image to consistent multi-view dif- fusion base model.arXiv preprint arXiv:2310.15110, 2023

    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. 3

    arXiv

  16. [24]

    Mvdream: Multi-view diffusion for 3d gen- eration.arXiv preprint arXiv:2308.16512, 2023

    Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d gen- eration.arXiv preprint arXiv:2308.16512, 2023. 2, 3

    2023 arXiv

  17. [25]

    Very deep convo- lutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556, 2014

    Karen Simonyan and Andrew Zisserman. Very deep convo- lutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556, 2014. 6

    2014 arXiv

  18. [26]

    Splatter image: Ultra-fast single-view 3d recon- struction

    Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d recon- struction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208– 10217, 2024. 3

    2024

  19. [27]

    Lgm: Large multi-view gaussian model for high-resolution 3d content creation

    Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. InEuropean Conference on Computer Vision, pages 1–18. Springer, 2025. 2, 6

    2025

  20. [28]

    Ref-nerf: Struc- tured view-dependent appearance for neural radiance fields

    Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Struc- tured view-dependent appearance for neural radiance fields. In2022 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 5481–5490. IEE...

    2022

  21. [29]

    Sv3d: Novel multi-view syn- thesis and 3d generation from a single image using latent video diffusion

    Vikram V oleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view syn- thesis and 3d generation from a single image using latent video diffusion. InEuropean Conference on Computer ...

    2025

  22. [30]

    Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation

    Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619–12629, 2023. 3

    2023

  23. [31]

    Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004. 6

    2004

  24. [32]

    Unique3d: High-quality and efficient 3d mesh generation from a single image.arXiv preprint arXiv:2405.20343, 2024

    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.arXiv preprint arXiv:2405.20343, 2024. 1, 2

    2024 arXiv

  25. [33]

    Reconfusion: 3d reconstruction with diffusion priors

    Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. InProceedings of the IEEE/CVF Conference on Computer Vision and Patte...

    2024

  26. [34]

    Diffusionerf: Regularizing neural radiance fields with denoising diffu- sion models

    Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffu- sion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4180– 4189, 2023. 1, 2, 3

    2023

  27. [35]

    Structured 3d latents for scalable and versatile 3d gen- eration, 2024

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d gen- eration, 2024. 2, 3, 6

    2024

  28. [36]

    Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models.arXiv preprint arXiv:2404.07191,

    Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models.arXiv preprint arXiv:2404.07191,

    arXiv

  29. [37]

    Grm: Large gaussian reconstruction model for ef- ficient 3d reconstruction and generation.arXiv preprint arXiv:2403.14621, 2024

    Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wet- zstein. Grm: Large gaussian reconstruction model for ef- ficient 3d reconstruction and generation.arXiv preprint arXiv:2403.14621, 2024. 2

    2024 arXiv

  30. [38]

    Gaussianob- ject: Just taking four images to get a high-quality 3d object with gaussian splatting.arXiv preprint arXiv:2402.10259,

    Chen Yang, Sikuang Li, Jiemin Fang, Ruofan Liang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Gaussianob- ject: Just taking four images to get a high-quality 3d object with gaussian splatting.arXiv preprint arXiv:2402.10259,

    arXiv

  31. [39]

    Lm-gaussian: Boost sparse-view 3d gaussian splatting with large model priors.arXiv preprint arXiv:2409.03456, 2024

    Hanyang Yu, Xiaoxiao Long, and Ping Tan. Lm-gaussian: Boost sparse-view 3d gaussian splatting with large model priors.arXiv preprint arXiv:2409.03456, 2024. 1, 2, 3

    2024 arXiv

  32. [40]

    Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis.arXiv preprint arXiv:2409.02048, 2024

    Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis.arXiv preprint arXiv:2409.02048, 2024. 3

    2024 arXiv

  33. [41]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 6

    2018

  34. [42]

    Cvt-xrf: Contrastive in-voxel transformer for 3d consis- tent radiance fields from sparse inputs

    Yingji Zhong, Lanqing Hong, Zhenguo Li, and Dan Xu. Cvt-xrf: Contrastive in-voxel transformer for 3d consis- tent radiance fields from sparse inputs. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21466–21475, 2024. 2

    2024

  35. [43]

    Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers

    Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern ...

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.