Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

ConsistentDreamer: View-Consistent Meshes Through Balanced Multi-View Gaussian Optimization

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 ConsistentDreamer generates view-consistent 3D meshes from a single image by using generated multi-view images as both ground truth for fine details and as the conditioning anchor for a diffusion prior that guides unseen views.

desk verdict A coherent engineering paper whose pipeline is genuinely plausible, but the headline view-consistency claim is only weakly supported by the reported metrics and is contingent on the unvalidated multi-view generator. read the letter →

arxiv 2502.09278 v3 pith:WUL2PNUC submitted 2025-02-13 cs.CV

Onat Şahin , Mohammad Altillawi , George Eskandar , Carlos Carbone , Ziyuan Liu This is my paper · ORCID
classification cs.CV
keywords image-to-3Dviewconsistencyscoredistillationsampling3DGaussiansplattingmeshextractionhomoscedasticuncertaintymulti-viewgeneration
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

ConsistentDreamer claims that the way to get view-consistent 3D meshes from a single image is to stop trusting the diffusion prior to generate every view independently. Instead, it first generates a small fixed set of multi-view images and uses those as a stable reference: a reconstruction loss pins the fine details of those views onto a 3D Gaussian representation, while a score-distillation loss uses the closest reference view to guide random views in between. The two optimizations are balanced automatically with weights learned from homoscedastic uncertainty, which the authors show stabilizes optimization and produces representations with fewer Gaussians. If correct, the method offers a practical pipeline that extracts a textured mesh from one photo in about a minute, with better view consistency and detail than existing image-to-3D systems.

What carries the argument

The central mechanisms are (1) closest-view SDS guidance, where the angular distance between the sampled view and the six prior views decides which generated image conditions the diffusion prior; (2) a per-iteration reconstruction loss that treats the generated multi-view images as ground truth; and (3) homoscedastic uncertainty weights, learned with backpropagation, that rescale the SDS and reconstruction terms as $\omega = \log\sigma^2$ so the optimization stays balanced. The opacity, depth-distortion, and normal-alignment losses are secondary machinery that turns the Gaussian blob into an extractable mesh.

What would settle it

Run ConsistentDreamer on a set of objects where Zero123++ produces visibly inconsistent multi-view images (for example, reflective or transparent objects on OmniObject3D) and measure the CLIP consistency of the final mesh with the input image. If the mesh stays consistent despite inconsistent prior views, the pipeline is doing the claimed work; if it inherits the generator's inconsistencies, the core premise fails.

Watch

Extended reading notes

Core claim

The central discovery is that the one-to-many ambiguity of image-to-3D is tamed by separating guidance from supervision. Six multi-view images generated in a first stage serve a dual role: they are pixel-level ground truth for the fine-detail reconstruction loss, and they are the conditioning set for the SDS loss, which always uses the reference image closest to the sampled viewpoint as the prompt. Because the closest-view angle difference is small, the diffusion prior makes more consistent predictions, and because the reconstruction loss runs every iteration, the details of all generated views are baked into the Gaussians. The view-consistency problem is thus attacked from both sides: rough geometry is pinned down by the SDS loss between views, and fine appearance is pinned by direct reconstruction at the six anchor views. The paper argues that the resulting meshes are more consistent in content and color across all views, and that the homoscedastic uncertainty weighting is what makes running both loss families simultaneously stable and efficient.

Load-bearing premise

The load-bearing premise is that the generated multi-view images are coherent and accurate enough to serve as both pixel-level ground truth and SDS conditioning even though the original input image is discarded during optimization, so any error in those generated views passes straight into the final mesh.

Editorial extensions

If this is right

  • If correct, image-to-3D pipelines can be built without relying on a single diffusion prior to guarantee consistency; generated multi-view anchors plus an SDS prior are enough.
  • The automatic uncertainty-based weighting removes the need for hand-tuned loss schedules, making the optimization more stable and the Gaussian representation leaner.
  • The method generalizes across multi-view generators (Zero123++, Wonder3D, ImageDream) and view counts, and ground-truth multi-view input turns it into a high-quality multi-view-to-3D reconstructor.
  • Surface-oriented losses (opacity, depth distortion, normal alignment) make Gaussian representations usable for mesh extraction, addressing a known bottleneck for embodied-AI asset pipelines.

Reading between the lines

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

  • The pipeline's ceiling is set by the multi-view generator: since the original image is discarded and the reconstruction loss forces exact matching to generated views, any systematic error in the first stage (e.g., hallucinations, missing parts, lighting inconsistencies) propagates directly into the mesh; a future version could retain the input view as an anchor or add a consistency filter.
  • The authors compare CLIP consistency with the input image, but the more natural test of 'view consistency' is the pairwise agreement between rendered views themselves; a stricter metric might reveal artifacts that CLIP-average hides.
  • Because the method works with any fixed set of reference views, it could be adapted to interactive or multi-image settings where the anchors are user-provided rather than generated, effectively unifying image-to-3D and multi-view reconstruction.
  • The ~1 minute runtime on a single 24GB GPU is attractive for simulation asset pipelines, but scaling to higher-resolution images or more anchor views may require rebalancing the uncertainty weights; the paper does not study that regime.
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 ConsistentDreamer, an image-to-3D pipeline that first generates six multi-view reference images with Zero123++, then optimizes a 3D Gaussian representation using an SDS loss conditioned on the closest reference view and simultaneous pixel-level reconstruction and mask losses against the same references. The two optimization tasks are balanced by homoscedastic-uncertainty weights that are updated during training, and additional opacity, depth-distortion, and normal-alignment losses are used to improve mesh extraction. The method is evaluated against seven baselines on GSO, OmniObject3D, and a subset of Objaverse, reporting SSIM, PSNR, LPIPS, CLIP similarity, and CLIP consistency, plus ablations on the loss components and on the choice of multi-view generator.

Significance. If the results hold, the main contribution is a practical demonstration that combining generated multi-view references for both SDS conditioning and pixel-level reconstruction yields meshes with competitive LPIPS and CLIP scores at a modest Gaussian count. The ablation study in Tables 4 and 5 is informative and supports the role of both the rough and fine optimizations, and the runtime of about one minute per mesh is attractive for embodied-AI asset generation. The central view-consistency claim is, however, conditional on the reliability of the external generator, and the reported CLIP consistency metric does not actually measure pairwise consistency among views; these issues need to be resolved before the headline claim is convincing.

major comments (4)
  1. [Section 5.1, Metrics] The CLIP Consistency metric is defined as the mean CLIP similarity between the input image and each of the 8 renders. This measures fidelity to the input view, not consistency among the generated views. The abstract and Section 6 claim 'better view consistency across views,' but no table reports pairwise agreement between renders at different azimuths, nor the variance of per-view features. Please add a direct pairwise view-consistency metric (e.g., CLIP or LPIPS similarity between all pairs of output views), or revise the claims to match what is actually measured.
  2. [Sections 4 and 5.1, Eqs. (5)-(7)] The optimization uses the six Zero123++ images as both pixel-level ground truth and SDS conditioning, while the original input image is explicitly discarded during optimization. This means any hallucination or inconsistency in the generated priors is inherited by the final mesh. Table 5 provides direct evidence for this dependence: replacing the generated priors with ground-truth GSO views raises PSNR from 19.06 to 21.35 and CLIP similarity from 0.90 to 0.94. The paper supplies no per-object analysis of Zero123++ failures, no consistency filter for bad prior views, and no discussion of how often such failures occur. To support the 'ensures better view consistency' claim, either add a robustness analysis or filtering mechanism, or explicitly qualify the claim as conditional on prior-view quality.
  3. [Tables 1-3, Section 5.2] The reported improvements over baselines are small (e.g., GSO LPIPS 0.13 vs 0.14 and CLIP consistency 0.87 vs 0.87; Objaverse LPIPS 0.15 vs 0.15) and no error bars or significance tests are given anywhere. Since the central quantitative claim rests on these margins, please report per-object standard deviations or confidence intervals and a paired statistical test (e.g., Wilcoxon signed-rank) over the 30 objects in each dataset. Without this, it is not clear which of the tabulated differences are meaningful.
  4. [Table 3 and Abstract] On Objaverse, the comparison is limited to DreamGaussian, HarmonyView, and Wonder3D, with the stated justification that the other baselines are trained or fine-tuned on Objaverse. The rationale is reasonable, but it means the abstract's unqualified 'state-of-the-art' claim is not supported on that dataset. Please qualify the claim to reflect the restricted comparison, or extend the evaluation with baselines whose training data are disclosed.
minor comments (5)
  1. [Eq. (8)] The opacity loss is written as -α log α - (1-α) log(1-α), which is the entropy and is minimized at both α=0 and α=1, not specifically at α=1 as claimed in the text ('drive the opacities ... to 1'). If the goal is to force opacities to 1, the loss should be -log α or a binary cross-entropy with target 1; if the goal is binarization, the text should say so.
  2. [Section 5.1] The sentence 'we only use the generated images and not the original input image during optimization' is a surprising design choice that deserves an explicit justification, particularly given that the generated images are also the only supervision for the fine-detail reconstruction.
  3. [Figure 7] The loss-weight plot would be easier to interpret if the curves were labeled with the actual values of ω_SDS and ω_rec, and if the y-axis scale were specified.
  4. [Section 5.1 and Tables 1-3] The evaluation renders 8 views at zero elevation, while the six prior viewpoints have elevations of 20° and -10°. Since the reconstruction loss directly supervises only the prior viewpoints, consider also reporting results at those viewpoints to separate the effect of the reconstruction loss from generalization to unseen elevations.
  5. [References] Reference [28] is listed as 'Sugar' but the correct name is 'SuGaR'; also check the text for inconsistent spacing in the word 'diffusion' that appears in several places.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pipeline is supervised by external multi-view priors and evaluated on external ground-truth datasets.

full rationale

The paper's central claim is an engineering claim about a pipeline, and its results do not reduce by construction to its inputs. The six multi-view prior images generated by Zero123++ are used as external supervision in two places: the reconstruction and mask losses (Eqs. 6-7) directly compare rendered views with these prior images, and the SDS loss (Eq. 5) conditions the Zero-1-to-3 prior on the closest prior image. These are choices of supervision signal, not definitions of the output: the optimized Gaussian parameters are produced by gradient descent on a composite loss, and the final mesh is evaluated against held-out ground-truth renders from GSO, OmniObject3D, and Objaverse. Table 5 shows that changing the generator or using ground-truth views changes quality, which confirms the supervision images are informative inputs rather than tautological targets. The homoscedastic uncertainty weights (Eqs. 9-10) are internal balancing coefficients learned on the optimization losses; they are not fitted to the evaluation metrics and are not relabeled as predictions. No load-bearing step is justified by a self-citation: the cited losses and priors are external and independently published, and the paper does not invoke a uniqueness theorem from its own authors. The CLIP consistency metric, defined as mean similarity between the input image and eight renders, does not directly measure pairwise agreement, but that is a metric-validity concern, not a circularity. Overall, the derivation chain is self-contained with respect to external benchmarks; the use of generated images as ground truth is a methodological dependency, not circular reasoning.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on two classes of assumptions: the reliability of external generative models (Zero123++ multi-view generation and Zero-1-to-3 SDS) that supply the supervision, and the transferability of existing loss formulations and uncertainty weighting to this pipeline. No new physical or architectural entities are introduced. The fixed loss coefficients and learned uncertainty weights are the main numeric choices; the viewpoint set is a hand-selected design choice not justified by an objective criterion.

free parameters (3)
  • homoscedastic uncertainty weights omega_SDS and omega_rec = learned per scene, initialized at 0, updated with Adam over 500 iterations
    Eq. 9-10: these scalars determine the balance between rough and fine optimization; they are not fixed hyperparameters but learned variables, so the balancing claim depends on their behavior.
  • fixed loss coefficients lambda_RGB, lambda_A, lambda_opacity, lambda_normal, lambda_depth, lambda_SDS = 10000, 1000, 0.1, 0.05, 100, 1
    Section 5.1: hand-set constants adopted from previous work; they define the loss balance and final mesh quality, and are not tuned on the evaluation set.
  • number and viewpoints of multi-view prior images = N=6; elevations 20 and -10 degrees, azimuths 30, 90, 150, 210, 270, 330 degrees
    Section 5.1: chosen by the authors without an objective criterion; the closest-view SDS and reconstruction coverage depend on this hand-picked camera layout, and Table 5 shows the viewpoint set affects results.
assumptions (5)
  • domain assumption Zero123++ multi-view images are accurate and view-consistent enough to act as pixel-level ground truth for the fine-detail reconstruction loss.
    Section 4, Eq. 6-7: reconstruction and mask losses compare rendered views against these generated images. If the generator produces a wrong or inconsistent view, that error is enforced directly into the mesh with no filtering.
  • domain assumption Zero-1-to-3 SDS gradients, conditioned on the closest generated prior view, improve the rough shape for every randomly sampled viewpoint.
    Section 4, Eq. 5: the method assumes the view-conditioned diffusion prior gives useful gradients over the whole camera sphere; this is empirically plausible but not proven.
  • domain assumption Homoscedastic uncertainty weighting from Cipolla et al. is valid for the SDS and reconstruction losses in this pipeline.
    Eq. 9-10 apply a multi-task learning weighting scheme to two losses of different scales; the paper does not verify the homoscedastic Gaussian noise assumption that justifies the derivation in the cited work.
  • domain assumption Rade-GS depth distortion and normal alignment losses transfer unchanged to this setting and improve mesh extraction.
    Section 4: opacity, depth distortion, and normal alignment losses are taken from the cited works with a pointer to a reference for details rather than re-derived; the interaction with SDS and reconstruction losses is not analyzed.
  • domain assumption The evaluation protocol (30 objects per dataset, 8 zero-elevation views, means without error bars) is sufficient to support the claim of better view consistency.
    Section 5.1-5.2: no seeds or variance are reported, and the views at elevations 20 and -10 degrees that are used during optimization are not quantitatively evaluated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ConsistentDreamer: View-Consistent Meshes Through Balanced Multi-View Gaussian Optimization." pith.science (2026). https://pith.science/paper/WUL2PNUC

@misc{pith2026250209278,
  author       = {Pith},
  title        = {Pith review of: ConsistentDreamer: View-Consistent Meshes Through Balanced Multi-View Gaussian Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WUL2PNUC}},
  note         = {Machine review of arXiv:2502.09278}
}
read the original abstract

Recent advances in diffusion models have significantly improved 3D generation, enabling the use of assets generated from an image for embodied AI simulations. However, the one-to-many nature of the image-to-3D problem limits their use due to inconsistent content and quality across views. Previous models optimize a 3D model by sampling views from a view-conditioned diffusion prior, but diffusion models cannot guarantee view consistency. Instead, we present ConsistentDreamer, where we first generate a set of fixed multi-view prior images and sample random views between them with another diffusion model through a score distillation sampling (SDS) loss. Thereby, we limit the discrepancies between the views guided by the SDS loss and ensure a consistent rough shape. In each iteration, we also use our generated multi-view prior images for fine-detail reconstruction. To balance between the rough shape and the fine-detail optimizations, we introduce dynamic task-dependent weights based on homoscedastic uncertainty, updated automatically in each iteration. Additionally, we employ opacity, depth distortion, and normal alignment losses to refine the surface for mesh extraction. Our method ensures better view consistency and visual quality compared to the state-of-the-art.

Figures

Figures reproduced from arXiv: 2502.09278 by the authors.

Figure 1
Figure 1. ConsistentDreamer for image-to-3D. Unlike prior optimization methods using only view-conditioned diffusion guidance or multi-view recon￾struction, we utilize both for balanced optimization of rough shape and fine details, improving view consistency in content and visual quality. ation speeds. However, the encoder-decoder structure leads to quality degradation from the input images, causing blurry tex￾tures and incon… view at source ↗
Figure 2
Figure 2. ConsistentDreamer pipeline. ConsistentDreamer is a Gaussian-based method for view-consistent 3D generation from a single image, guided by consistent multi-view images generated in a prior stage. Rough shape is optimized by improving random views with a diffusion conditioned on the closest prior view, while fine details are refined by comparing all prior views to corresponding views of the representation. A balance b… view at source ↗
Figure 3
Figure 3. Qualitative comparisons on image-to-3D mesh generation. We compare ConsistentDreamer against various methods using images from the internet. We include the input image and two diagonal views of each generated mesh from 45◦ and 225◦ azimuth angles. Our method gives the best results with consistent detail and color across views all around the object. Google Scanned Objects [31] (GSO) OmniObject3D [32] Objaverse [2] Gr… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Sample Meshes from our quantitative evaluation. We show some sample meshes from our quantitative evaluation shown on Tables 1, 2 and 3. We include the two diagonal views of each generated mesh from 45◦ and 225◦ azimuth angles, along with corresponding ground truth view…
Figure 7
Figure 7. Figure 7: The effect of loss balancing. We analyze the rough and fine losses throughout optimization without and with our proposed loss balancing in (a) and (b), respectively. (c) shows loss weights values over the course of the opti￾mization in (b). 5.3.2. Multi-View Generation…
Figure 6
Figure 6. Figure 6: Ablation study (qualitative). We qualitatively observe the effects of removing individual elements of our optimization on the final mesh surface. Rough Shape Optimization. When the multi-view SDS loss for rough shape optimization is removed, the optimization fails to f…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Hallo4D: Multi-Modal Hallucination Mitigation for Consistent Spatio-Temporal Generation

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Hallo4D uses vision-language models to detect and correct spatial and temporal mistakes in AI-generated 3D and 4D content, improving consistency without retraining the base generators.

Reference graph

Works this paper leans on

33 extracted references · 25 canonical work pages · cited by 1 Pith paper

  1. [1]

    Katara, Z

    P. Katara, Z. Xian, K. Fragkiadaki, Gen2sim: Scaling up robot learning in simulation with generative models, in: IEEE International Conference on Robotics and Automation, 2024, pp. 6672–6679

    work page 2024

  2. [2]

    Deitke, R

    M. Deitke, R. Liu, M. Wallingford, H. Ngo, O. Michel, A. Kusupati, A. Fan, C. Laforte, V . V oleti, S. Y . Gadre, E. VanderBilt, A. Kembhavi, C. V ondrick, G. Gkioxari, K. Ehsani, L. Schmidt, A. Farhadi, Objaverse- xl: A universe of 10m + 3d objects, arXiv preprint arXiv:2307.05663 (2023)

    arXiv 2023

  3. [3]

    Mildenhall, P

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

    2021

  4. [4]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, G. Drettakis, 3d gaussian splat- ting for real-time radiance field rendering, ACM Transactions on Graph- ics 42 (4) (July 2023)

    work page 2023

  5. [5]

    Poole, A

    B. Poole, A. Jain, J. T. Barron, B. Mildenhall, Dreamfusion: Text-to-3d using 2d diffusion, in: The Eleventh International Conference on Learn- ing Representations, 2023

    work page 2023

  6. [6]

    R. Liu, R. Wu, B. Van Hoorick, P. Tokmakov, S. Zakharov, C. V on- drick, Zero-1-to-3: Zero-shot one image to 3d object, in: Proceedings of the IEEE /CVF International Conference on Computer Vision, 2023, pp. 9298–9309

    work page 2023

  7. [7]

    J. Tang, J. Ren, H. Zhou, Z. Liu, G. Zeng, Dreamgaussian: Generative gaussian splatting for efficient 3d content creation, in: The Twelfth Inter- national Conference on Learning Representations, 2024

    work page 2024

  8. [8]

    G. Qian, J. Mai, A. Hamdi, J. Ren, A. Siarohin, B. Li, H.-Y . Lee, I. Sko- rokhodov, P. Wonka, S. Tulyakov, B. Ghanem, Magic123: One image to high-quality 3d object generation using both 2d and 3d di ffusion priors, in: The Twelfth International Conference on Learning Representations, 2024

    work page 2024

Show all 33 references
  1. [9]

    P. Wang, L. Liu, Y . Liu, C. Theobalt, T. Komura, W. Wang, Neus: Learn- ing neural implicit surfaces by volume rendering for multi-view recon- struction, NeurIPS (2021)

    2021

  2. [10]

    Y . Shi, P. Wang, J. Ye, L. Mai, K. Li, X. Yang, MVDream: Multi-view diffusion for 3d generation, in: The Twelfth International Conference on Learning Representations, 2024

    2024

  3. [11]

    P. Wang, Y . Shi, Imagedream: Image-prompt multi-view diffusion for 3d generation, arXiv preprint arXiv:2312.02201 (2023)

    2023 arXiv

  4. [12]

    R. Shi, H. Chen, Z. Zhang, M. Liu, C. Xu, X. Wei, L. Chen, C. Zeng, H. Su, Zero123++: a single image to consistent multi-view diffusion base model, CoRR abs/2310.15110 (2023)

    2023 arXiv

  5. [13]

    Y . Liu, C. Lin, Z. Zeng, X. Long, L. Liu, T. Komura, W. Wang, Sync- dreamer: Generating multiview-consistent images from a single-view im- age, in: The Twelfth International Conference on Learning Representa- tions, 2024

    2024

  6. [14]

    S. Woo, B. Park, H. Go, J.-Y . Kim, C. Kim, Harmonyview: Harmo- nizing consistency and diversity in one-image-to-3d, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 10574–10584

    2024

  7. [15]

    Long, Y .-C

    X. Long, Y .-C. Guo, C. Lin, Y . Liu, Z. Dou, L. Liu, Y . Ma, S.-H. Zhang, M. Habermann, C. Theobalt, W. Wang, Wonder3d: Single image to 3d us- ing cross-domain diffusion, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 9970–9980

    2024

  8. [16]

    Y . Hong, K. Zhang, J. Gu, S. Bi, Y . Zhou, D. Liu, F. Liu, K. Sunkavalli, T. Bui, H. Tan, LRM: Large reconstruction model for single image to 3d, in: The Twelfth International Conference on Learning Representations, 2024

    2024

  9. [17]

    J. Tang, Z. Chen, X. Chen, T. Wang, G. Zeng, Z. Liu, Lgm: Large multi-view gaussian model for high-resolution 3d content creation, arXiv preprint arXiv:2402.05054 (2024)

    2024 arXiv

  10. [18]

    Z. Wang, Y . Wang, Y . Chen, C. Xiang, S. Chen, D. Yu, C. Li, H. Su, J. Zhu, Crm: Single image to 3d textured mesh with convolutional recon- struction model, in: European Conference on Computer Vision, 2024

    2024

  11. [19]

    Tochilkin, D

    D. Tochilkin, D. Pankratz, Z. Liu, Z. Huang, A. Letts, Y . Li, D. Liang, C. Laforte, V . Jampani, Y .-P. Cao, Triposr: Fast 3d object reconstruction from a single image (2024). arXiv:2403.02151

    2024 arXiv

  12. [20]

    Z.-X. Zou, Z. Yu, Y .-C. Guo, Y . Li, D. Liang, Y .-P. Cao, S.-H. Zhang, Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers, in: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2024, pp...

    2024

  13. [21]

    A. Yu, V . Ye, M. Tancik, A. Kanazawa, pixelnerf: Neural radiance fields from one or few images, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 4578–4587

    2021

  14. [22]

    Y . Liu, S. Peng, L. Liu, Q. Wang, P. Wang, C. Theobalt, X. Zhou, W. Wang, Neural rays for occlusion-aware image-based rendering, in: IEEE/CVF Conference on Computer Vision and Pattern Recognition, IEEE Computer Society, Los Alamitos, CA, USA, 2022, pp. 7814–7823

    2022

  15. [23]

    Trevithick, B

    A. Trevithick, B. Yang, Grf: Learning a general radiance field for 3d rep- resentation and rendering, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 15182–15192

    2021

  16. [24]

    Zimny, J

    D. Zimny, J. Waczy ´nska, T. Trzci´nski, P. Spurek, Points2nerf: Generating neural radiance fields from 3d point cloud, Pattern Recognition Letters 185 (2024) 8–14

    2024

  17. [25]

    K. Jin, X. Gu, Z. Wang, Z. Kuang, Z. Wu, M. Tan, J. Yu, Semantic-aware hyper-space deformable neural radiance fields for facial avatar reconstruc- tion, Pattern Recognition Letters 185 (2024) 160–166

    2024

  18. [26]

    J. Tang, T. Wang, B. Zhang, T. Zhang, R. Yi, L. Ma, D. Chen, Make- it-3d: High-fidelity 3d creation from a single image with di ffusion prior, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 22819–22829

    2023

  19. [27]

    H. Chen, C. Li, G. H. Lee, Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance, ArXiv abs/2312.00846 (2023)

    2023 arXiv

  20. [28]

    Gu ´edon, V

    A. Gu ´edon, V . Lepetit, Sugar: Surface-aligned gaussian splatting for ef- ficient 3d mesh reconstruction and high-quality mesh rendering, in: Pro- ceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 5354–5363

    2024

  21. [29]

    Zhang, C

    B. Zhang, C. Fang, R. Shrestha, Y . Liang, X. Long, P. Tan, Rade-gs: Rasterizing depth in gaussian splatting (2024). arXiv:2406.01467

    2024 arXiv

  22. [30]

    Cipolla, Y

    R. Cipolla, Y . Gal, A. Kendall, Multi-task learning using uncertainty to weigh losses for scene geometry and semantics, in: IEEE /CVF Confer- ence on Computer Vision and Pattern Recognition, 2018, pp. 7482–7491

    2018

  23. [31]

    Downs, A

    L. Downs, A. Francis, N. Koenig, B. Kinman, R. M. Hickman, K. Rey- mann, T. B. McHugh, V . Vanhoucke, Google scanned objects: A high- quality dataset of 3d scanned household items, 2022 International Con- ference on Robotics and Automation (2022) 2553–2560

    2022

  24. [32]

    T. Wu, J. Zhang, X. Fu, Y . Wang, L. P. Jiawei Ren, W. Wu, L. Yang, J. Wang, C. Qian, D. Lin, Z. Liu, Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation, in: IEEE/CVF Conference on Computer Vision and Pattern Recognition (C...

    2023

  25. [33]

    T. Shen, J. Munkberg, J. Hasselgren, K. Yin, Z. Wang, W. Chen, Z. Goj- cic, S. Fidler, N. Sharp, J. Gao, Flexible isosurface extraction for gradient- based mesh optimization, ACM Trans. Graph. 42 (4) (jul 2023). 8

    2023

Pith tools

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