Pith. sign in

REVIEW 3 major objections 6 minor 30 references

Geometric Neural Process Fields

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read This paper claims that adding learned Gaussian geometric bases and hierarchical global-to-local latent variables to Neural Process Fields improves generalization for neural fields, giving state-of-the-art PSNR for 1-view and 2-view novel…

desk verdict A novel architecture for neural process fields with geometric bases and hierarchical latents, but the training objective contradicts the derived ELBO and the basis KL term is undefined. read the letter →

arxiv 2502.02338 v1 pith:AHO2TKUT submitted 2025-02-04 cs.CV cs.LG

Wenzhe Yin , Zehao Xiao , Jiayi Shen , Yunlu Chen , Cees G. M. Snoek , Jan-Jakob Sonke , Efstratios Gavves This is my paper · ORCID
classification cs.CVcs.LG
keywords neuralfieldsprocessesnovelviewsynthesisuncertaintyquantificationgeometricbaseshierarchicallatentvariablesradiancefew-shotgeneralization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that neural fields can generalize to new 3D scenes from one or two context images if the field is constrained by two structural inductive biases: a set of learned Gaussian bases that encode where scene information lives in space, and a hierarchy of global and per-ray latent variables that capture object-level and coordinate-specific uncertainty. It argues that previous probabilistic neural-field models, which condition directly on raw context pixels, ignore scene geometry and therefore waste limited observations; G-NPF instead routes all context through the geometric bases and latent variables. If the claim is right, the recipe is reusable: the same probabilistic model with Gaussian bases and hierarchical latents improves 2D image regression and 1D signal regression as well as 3D rendering, and it does so with fewer parameters than one of its main baselines. The payoff is a practical way to do few-shot novel view synthesis with calibrated uncertainty, which matters whenever only a handful of views of an object are available.

What carries the argument

The central machinery is the geometric basis set $\mathcal{B}_C = \{(\mathcal{N}(\mu_r, \Sigma_r), \omega_r)\}_{r=1}^R$, produced by a transformer encoder from the context set, together with the target-location representation $\langle x_T, \mathcal{B}_C \rangle = \mathrm{MLP}[\sum_r \exp(-\frac{1}{2}(x_T - \mu_r)^\top \Sigma_r^{-1} (x_T - \mu_r))\, \omega_r]$ of Eq. (6). Each basis pairs a Gaussian in coordinate space with a learned semantic embedding, so it acts as an information bottleneck that encodes where scene structure is and what it looks like. On top of these bases the model defines a hierarchical prior: a global latent $z_g$ aggregates all target-location representations to encode the whole object or scene, and per-ray (in 3D) or per-point (in 1D and 2D) local latents $z_l$ combine with $z_g$ through a transformer to encode coordinate-specific detail. The training objective is a hierarchical ELBO plus a KL term aligning context-derived bases $\mathcal{B}_C$ with target-derived bases $\mathcal{B}_T$; the posterior is inferred from target data during training and the prior from context data at test time.

What would settle it

Run G-NPF and VNP on the NeRF Synthetic 'materials' scene, or any object with strong specular, view-dependent reflections, with one context view, using the same encoder, NeRF decoder, and training schedule as the paper; if the PSNR gap over VNP shrinks or reverses because the R Gaussian bases cannot represent high-frequency radiance, the bottleneck premise is falsified. A cheaper check: set the base-alignment weight beta to 0 and see whether the improvement over VNP persists, which tests whether target-base supervision is load-bearing.

Watch

Extended reading notes

Core claim

The core discovery is that compressing context observations through a learned set of R Gaussian geometric bases, each basis a Gaussian with mean, covariance, and a semantic embedding, and then inferring a global latent variable plus ray- or point-local latent variables from those bases, yields better generalization for Neural Process Fields than conditioning on the raw context set. The paper derives a hierarchical ELBO in which the priors are supervised by target-derived bases available only at training, and a KL term aligns context bases with target bases. Across ShapeNet categories and NeRF Synthetic scenes, the model reports the highest PSNR among the compared methods for 1-view and 2-view novel view synthesis, and it also outperforms its probabilistic baseline VNP on 1D GP regression and its deterministic baselines on CelebA and Imagenette image regression. The same design also produces cross-category generalization without retraining, which the paper attributes to the bases and hierarchical latents carrying structure rather than memorized category appearance.

Load-bearing premise

The load-bearing premise is that a small set of Gaussian bases in coordinate space, plus one global and one per-ray latent variable, retains enough scene structure and view-dependent appearance for novel-view synthesis from one or two context images; if that Gaussian bottleneck discards high-frequency or view-dependent information, the reported gains would collapse.

Editorial extensions

If this is right

  • Any Neural-Process-based neural field can likely be upgraded by inserting Gaussian geometric bases and global-to-local hierarchical latents; the paper demonstrates this for 1D, 2D, and 3D signals with the same code path.
  • One-view and two-view novel view synthesis becomes practical: on ShapeNet, G-NPF's 1-view PSNR is 0.5-1.2 higher than VNP's depending on category, and on NeRF Synthetic it reports about 0.9 PSNR higher than GeFu at 3 views.
  • Because the model outputs a distribution over latent codes rather than a point estimate, it can flag which rays or regions are uncertain, a direct consequence of the probabilistic formulation.
  • The model generalizes across categories without retraining (trained on drums, evaluated on lego), implying the learned bases and latents capture transferable structure rather than category-specific appearance.
  • The approach is parameter-efficient: on ShapeNet cars it uses 24.0M parameters versus 34.3M for VNP while reporting higher PSNR.

Reading between the lines

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

  • The Gaussian bases can be read as a learned spatial dictionary that separates 'where' (center and covariance) from 'what' (semantic embedding), so the recipe may transfer to other coordinate-based decoders, such as signed-distance or material fields, whenever structure is mostly geometric; the paper only demonstrates radiance, image, and 1D signals.
  • A natural untested use of the predicted latents is active view selection: the variance of the ray-specific latent gives a per-ray uncertainty that could rank which next camera to acquire; nothing in the paper evaluates this.
  • Because the bases use the same scale-plus-quaternion covariance parameterization as 3D Gaussian splatting, the probabilistic hierarchy could be grafted onto splatting pipelines as a prior over primitive parameters; the paper only remarks on complementarity.
  • The position-only Gaussian bottleneck suggests a testable extension: letting bases depend on viewing direction as well as position would likely improve view-dependent appearance, which is the paper's least-examined failure mode.
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 Geometric Neural Process Fields (G-NPF), a probabilistic conditional neural field for few-shot regression and novel view synthesis. The method encodes context observations into a set of learned Gaussian geometric bases and uses hierarchical latent variables (one global, plus per-point or per-ray local) to modulate a shared MLP. Training is formulated via variational inference, with an additional KL term aligning context-derived bases with target-derived bases. Experiments cover 1D signal regression, 2D image regression, and 3D novel view synthesis on ShapeNet and NeRF Synthetic, reporting improved PSNR/SSIM/LPIPS over several baselines. The paper also includes ablations of the geometric bases and hierarchical latents.

Significance. If the central claims hold, G-NPF offers a reusable recipe for injecting geometric and hierarchical structure into probabilistic neural fields, with potential value across 1D, 2D, and 3D signals. The derivation in Appendix B is structurally sound as a hierarchical ELBO, and the paper provides broad empirical coverage plus a favorable parameter count relative to VNP on the ShapeNet Car benchmark. However, the significance is currently conditional: the training objective stated in the main text does not match the derived ELBO, the basis-alignment KL is undefined, and the headline empirical claims are presented without error bars or significance testing. These issues must be resolved before the claimed state-of-the-art results can be attributed to the proposed probabilistic mechanism.

major comments (3)
  1. [§3.5, Eq. (9); Appendix B, Eq. (15); Appendix F.4, Eq. (30)] The training objective is not the ELBO derived in the paper. The derived ELBO (Eq. 15 and Eq. 31) contains D_KL(q(zg|xT,BT) || p(zg|xT,BC)) and D_KL(q(zl|zg,xT,BT) || p(zl|zg,xT,BC)), with the variational posterior as the left argument, while Eq. (9) and Eq. (30) use the reversed ordering D_KL(p(zg|BC) || q(zg|BT)) and D_KL(p(zl|zg,BC) || q(zl|zg,BT)). Reversed KL divergences are not equal in general, so the optimized loss is a different functional from the variational bound stated in Eq. (15). Since the paper credits the probabilistic formulation for the reported gains, this mismatch is load-bearing and must be corrected or explicitly justified.
  2. [§3.5, Eq. (9); §3.2; Appendix F.4] The term D_KL(BC || BT) is undefined as written. BC and BT are unordered sets of Gaussian bases {N(mu_r, Sigma_r), omega_r}, and no standard KL divergence exists between two such sets. An index-matched computation would be permutation-sensitive, and a Gaussian-mixture KL has no closed form; neither construction is specified anywhere in the manuscript. Because this term is the only stated mechanism for transferring target-derived bases to context-only inference, the paper cannot be reproduced and the claimed supervision of the context prior by target bases is not well defined.
  3. [§4.1, Tables 1–3; §4.2] The central empirical claims are reported without error bars or significance tests. Tables 2 and 3 list single PSNR/SSIM/LPIPS values, and Table 1 gives no variance either, yet the text states that G-NPF 'outperforms baselines consistently' and 'by a significant margin'. In particular, Table 3 reports a 1-view GNT PSNR of 10.25 versus 20.07 for G-NPF, which suggests a collapsed baseline rather than a meaningful comparison; multiple seeds and standard deviations are needed before a state-of-the-art claim can be assessed.
minor comments (6)
  1. [Table 2; Table 3] Both tables are captioned 'Qualitative Comparison' but report quantitative PSNR, SSIM, and LPIPS values; the captions should say 'Quantitative Comparison'.
  2. [§3.5, Eq. (9)] With alpha and beta both set to 0.001 (Appendix D), the KL terms are three orders of magnitude smaller than the squared-error term in a typical loss scale; the paper should report whether the reported gains are robust to the choice of alpha and beta, since the current values make the variational terms nearly negligible.
  3. [§3.4, Eq. (6)–(8)] The notation is inconsistent between the main text (zg, zl,m) and Appendix F (zo, zr), and Eq. (27)–(28) reuse the same symbols with different meanings; unifying the notation would improve readability.
  4. [§3.3, Eq. (4)] The integral in Eq. (4) has unbalanced parentheses and writes p(zl|zg,xT,BC) dzl ... p(zg|xT,BC) dzg without a clear joint integration domain; the factorization in Eq. (5) is clearer and the paper should use it consistently.
  5. [§C.3] The sentence 'We use modulation to The latent variables' is incomplete and should be rephrased; the intended meaning is that the latent variables are used to modulate the MLP weights as described in Eq. (19)–(20).
  6. [§4.4, Table 6] The sensitivity analysis reports that more bases always improve PSNR, but without error bars or a computation-time tradeoff it is unclear whether the largest configuration overfits or is unstable; reporting runs at several base counts with standard deviations would strengthen this ablation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the prediction is trained with target-supervised variational losses and evaluated on held-out views; the Eq. 9/30 loss-formula inconsistency is a specification issue, not a circular reduction.

full rationale

The central derivation is self-contained and not circular. G-NPF predicts target signals from context-derived geometric bases BC and hierarchical latents; target-derived bases BT appear only in the training objective as variational supervision (Eq. 9, Appendix F.4), and at test time the predictive distribution p(yT|xT,BC) depends only on context. The hierarchical ELBO in Appendix B follows the standard Jensen's-inequality derivation, and the empirical objective in Eq. 9/30, though it reverses the KL orientation and leaves DKL(BC||BT) underspecified, is a reproducibility/correctness concern rather than a circular step: no equation reduces the prediction to its own input by construction. Reported gains are validated against external baselines (VNP, TransINR, GeFu, GNT, etc.) on held-out views. Self-citations to prior work by the same group (Papa et al. 2024; de Vries et al. 2023) are background references and are not load-bearing for the method's claims. No self-definitional, fitted-input-renamed-as-prediction, or self-citation-chain pattern is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

The central claim rests mostly on a learned encoder and variational inference, so the ledger is dominated by modeling choices rather than physical axioms. The Gaussian basis bottleneck and conditional independence factorization are assumptions that could be wrong for correlated, high-frequency signals. They are internal to the method and have no independent empirical support beyond the reported benchmarks.

free parameters (3)
  • number of geometric bases R = R=49, 169, or 484 for image regression; R=100 or 250 for NeRF (Table 6)
    Selected per task based on accuracy and computation. Table 6 shows more bases always improve PSNR, so the reported best results depend on a performance-driven choice.
  • KL weights alpha and beta = alpha=0.001, beta=0.001
    Set by hand without sensitivity analysis in Appendix D. They directly weight the KL terms in the loss in Eq. (9) and Eq. (30).
  • semantic embedding dimension dB = 32
    Chosen for the basis representation in Appendix C.1; no ablation is provided.
assumptions (4)
  • domain assumption Target outputs are conditionally independent given the hierarchical latent variables zg and zl (Eq. 5).
    Assumed in the factorization of the predictive distribution. If residual correlations between pixels or camera rays matter, the likelihood is misspecified.
  • domain assumption A mixture of R Gaussians weighted by learned semantic vectors, combined through a Gaussian radial basis interaction (Eq. 6), captures the spatial structure needed for generalization.
    Core inductive bias of the paper. No analysis is given of how much information is lost by the Gaussian bottleneck.
  • standard math The volume rendering integral in Eq. (10) is the correct observation model for NeRF training.
    Standard NeRF assumption borrowed from classical volume rendering; not specific to this paper.
  • standard math Jensen's inequality and the factorized variational posterior in Appendix B produce a valid ELBO.
    Standard variational inference; the derivation appears structurally correct, with only notation issues.
invented entities (2)
  • Geometric bases BC = {(N(mu_r, Sigma_r), omega_r)}
    purpose: Encode sparse context into a small set of Gaussian spatial priors with semantic vectors, creating the bottleneck and spatial inductive bias used by all downstream latent variables.
    No falsifiable prediction outside the paper. It is a learned internal representation evaluated only through downstream PSNR.
  • Hierarchical latent variables zg and zl (object-level and point/ray-level)
    purpose: Carry global and local scene uncertainty into the modulated NeRF MLP.
    Latent variables are standard model components; here they are not tied to any externally measurable quantity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Geometric Neural Process Fields." pith.science (2026). https://pith.science/paper/AHO2TKUT

@misc{pith2026250202338,
  author       = {Pith},
  title        = {Pith review of: Geometric Neural Process Fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AHO2TKUT}},
  note         = {Machine review of arXiv:2502.02338}
}
read the original abstract

This paper addresses the challenge of Neural Field (NeF) generalization, where models must efficiently adapt to new signals given only a few observations. To tackle this, we propose Geometric Neural Process Fields (G-NPF), a probabilistic framework for neural radiance fields that explicitly captures uncertainty. We formulate NeF generalization as a probabilistic problem, enabling direct inference of NeF function distributions from limited context observations. To incorporate structural inductive biases, we introduce a set of geometric bases that encode spatial structure and facilitate the inference of NeF function distributions. Building on these bases, we design a hierarchical latent variable model, allowing G-NPF to integrate structural information across multiple spatial levels and effectively parameterize INR functions. This hierarchical approach improves generalization to novel scenes and unseen signals. Experiments on novel-view synthesis for 3D scenes, as well as 2D image and 1D signal regression, demonstrate the effectiveness of our method in capturing uncertainty and leveraging structural information for improved generalization.

Figures

Figures reproduced from arXiv: 2502.02338 by the authors.

Figure 1
Figure 1. Illustration of the proposed G-NPF. ditional Neural Fields offer an interesting path forward, they still suffer from the unconstrained nature of stochas￾tic gradient descent and the over-parameterized nature of neural networks (Papa et al., 2024), thus making representa￾tion learning and generalization to few-shot settings hard, whether for 1-D (e.g., for time series), 2-D (e.g., for PINNs), or 3-D (e.g., for occupa… view at source ↗
Figure 2
Figure 2. Graphical model for the proposed geometric neural processes. C directly. Here R is the number of bases. These geometric bases must create an information bottleneck through which we embed structure to the context set C, thus R ≪ ∥C∥ = N. Each geometric basis br =  N (µr, Σr), ωr  contains a Gaussian distribution N in the 2D spatial plane with covari￾ance Σr, centered around a 2D coordinate µr. Note that when extend… view at source ↗
Figure 3
Figure 3. Visualizations of image regression results on CelebA (left) and Imagenette (right). ing (162,000 images), validation (20,000 images), and test (20,000 images) sets. Imagenette dataset (Howard, 2020). Imagenette is a cu￾rated subset comprising 10 classes from the 1,000 classes in ImageNet (Deng et al., 2009), consists of roughly 9,000 training images and 4,000 testing images. Quantitative results. We give quantitativ… view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Qualitative results of the proposed G-NPF on novel view synthesis of ShapeNet objects. Both 1-view (top) and 2-view (bottom) context results are presented [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Using transformer encoder to generate ray-specific latent variable zl . C.2. Hierarchical Latent Variables At the object level, the distribution of the global latent variable zg is obtained by aggregating all location representations from (BC , xT ). We assume that p(z…
Figure 6
Figure 6. Figure 6: More image regression results on the Imagenette dataset. Left: ground truth; Right: prediction. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Image completion visualization on CelebA using 10% (left) and 20% (right) context. E. More Experimental Results E.1. Image Regression We provide more image regression results to demonstrate the effectiveness of our method as shown in [PITH_FULL_IMAGE:figures/full_fig_…
Figure 8
Figure 8. Figure 8: Qualitative comparison with GNT on 1-view setting. GeomNP GT GNT [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Qualitative comparison of cross-category ability. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: More NeRF results on novel view synthesis task on ShapeNet objects. Comparison with VNP Context VNP Ours GT Context VNP Ours GT [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Comparison between the proposed method and VNP on novel view synthesis task for ShapeNet objects. Our method has a better rendering quality than VNP for novel views. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: Training time vs. PSNR on the ShapeNet Car dataset. Our method (GeomNP) demonstrates faster convergence and higher final PSNR compared to the baseline (VNP). Context w/o global w/o local GeomNP GT [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]
Figure 13
Figure 13. Figure 13: Qualitative ablation of the hierarchical latent variables (global and local variables). 20 [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 14
Figure 14. Figure 14: Qualitative comparisons of Multi-view results on the Drums class of the NeRF synthetic dataset. F. Extending G-NPF to NeRFs Notations. We denote 3D world coordinates by p = (x, y, z) and a camera viewing direction by d = (θ, ϕ). Each point in 3D space have its color c…
Figure 15
Figure 15. Figure 15: Complete rendering from 3D points to a 2D pixel. Background on Neural Radiance Fields. We formally describe Neural Radiance Field (NeRF) (Mildenhall et al., 2021; 21 [PITH_FULL_IMAGE:figures/full_fig_p021_15.png]
Figure 16
Figure 16. Figure 16: Illustration of our Geometric Neural Processes. We cast radiance field generalization as a probabilistic modeling problem. Specifically, we first construct geometric bases BC in 3D space from the 2D context sets Xe C , Ye C to model the 3D NeRF function (Section F.2).…
Figure 17
Figure 17. Figure 17: Graphical model for the proposed geometric neural processes. Within the hierarchical Bayes framework, zo encodes the entire object information from all target inputs and the geometric bases {XT , BC } in the global level; while every z n r encodes ray-specific informa…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 10 canonical work pages

  1. [1]

    NeRF in detail: Learning to sample for view synthesis

    Arandjelovi´c, R. and Zisserman, A. Nerf in detail: Learning to sample for view synthesis. arXiv preprint arXiv:2106.05264,

    work page Pith review arXiv

  2. [4]

    For NeRF-related experiments, we follow the baselines (Chen & Wang, 2022; Guo et al.,

    work page 2022

  3. [5]

    Imagenet: A large-scale hierarchical image database

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

    2009

  4. [7]

    From data to functa: Your data point is a function and you can treat it like one

    Dupont, E., Kim, H., Eslami, S., Rezende, D., and Rosen- baum, D. From data to functa: Your data point is a function and you can treat it like one. arXiv preprint arXiv:2201.12204, 2022a. Dupont, E., Loya, H., Alizadeh, M., Goli´nski, A., Teh, Y . W., and Doucet, A. Coin++: Neural compression across modalities. arXiv preprint arXiv:2201.12904, 2022b. Erko...

    arXiv

  5. [10]

    Versatile Neural Processes for Learning Implicit Neural Representations

    Guo, Z., Lan, C., Zhang, Z., Lu, Y ., and Chen, Z. Versatile neural processes for learning implicit neural representa- tions. arXiv preprint arXiv:2301.08883,

    work page Pith review arXiv

  6. [11]

    M., and Le, Q

    Ha, D., Dai, A. M., and Le, Q. V . Hypernetworks. ArXiv, abs/1609.09106,

    arXiv

  7. [12]

    Lrm: Large recon- struction model for single image to 3d

    Hong, Y ., Zhang, K., Gu, J., Bi, S., Zhou, Y ., Liu, D., Liu, F., Sunkavalli, K., Bui, T., and Tan, H. Lrm: Large recon- struction model for single image to 3d. arXiv preprint arXiv:2311.04400,

    arXiv

  8. [13]

    Kim, H., Mnih, A., Schwarz, J., Garnelo, M., Eslami, A., Rosenbaum, D., Vinyals, O., and Teh, Y . W. Attentive neural processes. arXiv preprint arXiv:1901.05761,

    arXiv 1901

Show all 30 references
  1. [14]

    GeomNP GT GNT Figure 14: Qualitative comparisons of Multi-view results on the Drums class of the NeRF synthetic dataset. F. Extending G-NPF to NeRFs Notations. We denote 3D world coordinates by p = (x, y, z) and a camera viewing direction by d = (θ, ϕ). Each point in 3D space ...

    2021

  2. [15]

    Kingma, D. P. and Welling, M. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114,

    arXiv

  3. [16]

    R., Strathmann, H., Zoran, D., Schneider, R

    Moreno, P., Kosiorek, A. R., Strathmann, H., Zoran, D., Schneider, R. G., Winckler, B., Markeeva, L., We- ber, T., and Rezende, D. J. Laser: Latent set repre- sentations for 3d generative modeling. arXiv preprint arXiv:2301.05747,

    arXiv

  4. [17]

    and Grover, A

    Nguyen, T. and Grover, A. Transformer neural processes: Uncertainty-aware meta learning via sequence modeling. arXiv preprint arXiv:2207.04179,

    arXiv

  5. [18]

    On first-order meta-learning algorithms

    Nichol, A., Achiam, J., and Schulman, J. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999,

    arXiv

  6. [19]

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

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

    arXiv

  7. [20]

    Learning robust generaliz- able radiance field with visibility and feature augmented point representation

    Wang, J., Zhang, Z., and Xu, R. Learning robust generaliz- able radiance field with visibility and feature augmented point representation. arXiv preprint arXiv:2401.14354,

    arXiv

  8. [22]

    Fg-nerf: Flow-gan based probabilistic neural radi- ance field for independence-assumption-free uncertainty estimation

    Wei, S., Zhang, J., Wang, Y ., Xiang, F., Su, H., and Wang, H. Fg-nerf: Flow-gan based probabilistic neural radi- ance field for independence-assumption-free uncertainty estimation. arXiv preprint arXiv:2309.16364,

    arXiv

  9. [23]

    Continuous pde dynamics forecast- ing with implicit neural representations

    Yin, Y ., Kirchmeyer, M., Franceschi, J.-Y ., Rakotomamonjy, A., and Gallinari, P. Continuous pde dynamics forecast- ing with implicit neural representations. arXiv preprint arXiv:2209.14855,

    arXiv

  10. [24]

    Neural Radiance Field Rendering In this section, we outline the rendering function of NeRF (Mildenhall et al., 2021)

    12 Geometric Neural Process Fields A. Neural Radiance Field Rendering In this section, we outline the rendering function of NeRF (Mildenhall et al., 2021). A 5D neural radiance field represents a scene by specifying the volume density and the directional radiance emitted at ev...

    2021

  11. [28]

    as coordinates to color and density mappings. Learning a NeRF function of a 3D object is an inverse problem where we only have indirect observations of arbitrary 2D views of the 3D object, and we want to infer the entire 3D object’s geometry and appearance. With the NeRF funct...

    1984

  12. [29]

    overfitting

    Details about the integration are given in Appendix A. F.1. Probabilistic NeRF Generalization Deterministic Neural Radiance Fields Neural Radiance Fields are normally considered as an optimization routine in a deterministic setting (Mildenhall et al., 2021; Barron et al., 2021...

    2021

  13. [30]

    We first represent each target location by integrating the geometric bases, i.e., < xn T , BC >, which aggregates the relevant locality and semantic information for the given input

    In the modeling of GeomNP, the prior distribution of each hierarchical latent variable is conditioned on the geometric bases and target input. We first represent each target location by integrating the geometric bases, i.e., < xn T , BC >, which aggregates the relevant localit...

    2023

  14. [32]

    We use a self-attention module, denoted Att, to extract these Gaussian parameters from the context data

    Self-Attention Construction. We use a self-attention module, denoted Att, to extract these Gaussian parameters from the context data. Concretely, µi, Σi = Att xC, yC , ω i = Att xC, yC , (17) where each call to Att produces M tokens of hidden dimension D. An MLP then maps each...

    2023

  15. [2009]

    W., and Doucet, A

    9 Geometric Neural Process Fields Dupont, E., Teh, Y . W., and Doucet, A. Generative models as distributions of functions. arXiv preprint arXiv:2102.04776,

    arXiv

  16. [2017]

    W., Rezende, D., and Eslami, S

    Garnelo, M., Rosenbaum, D., Maddison, C., Ramalho, T., Saxton, D., Shanahan, M., Teh, Y . W., Rezende, D., and Eslami, S. A. Conditional neural processes. In Interna- tional conference on machine learning, pp. 1704–1713. PMLR, 2018a. Garnelo, M., Schwarz, J., Rosenbaum, D., Vi...

    arXiv

  17. [2019]

    Neural processes with stochastic attention: Paying more attention to the context dataset

    Kim, M., Go, K., and Yun, S.-Y . Neural processes with stochastic attention: Paying more attention to the context dataset. arXiv preprint arXiv:2204.05449,

    arXiv

  18. [2020]

    P., Foong, A

    Gordon, J., Bruinsma, W. P., Foong, A. Y ., Requeima, J., Dubois, Y ., and Turner, R. E. Convolutional conditional neural processes. arXiv preprint arXiv:1910.13556,

    1910 arXiv

  19. [2021]

    P., Markou, S., Requiema, J., Foong, A

    Bruinsma, W. P., Markou, S., Requiema, J., Foong, A. Y ., Andersson, T. R., Vaughan, A., Buonomo, A., Hosking, J. S., and Turner, R. E. Autoregressive conditional neural processes. arXiv preprint arXiv:2303.14468,

    arXiv

  20. [2022]

    Explicit correspondence matching for generalizable neural radiance fields

    Chen, Y ., Xu, H., Wu, Q., Zheng, C., Cham, T.-J., and Cai, J. Explicit correspondence matching for generalizable neural radiance fields. arXiv preprint arXiv:2304.12294, 2023a. Chen, Y ., Xu, H., Zheng, C., Zhuang, B., Pollefeys, M., Geiger, A., Cham, T.-J., and Cai, J. Mvspl...

    arXiv

  21. [2023]

    X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al

    Chang, A. X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012,

    arXiv

  22. [2024]

    Is attention all that nerf needs? arXiv preprint arXiv:2207.13298,

    Wang, P., Chen, X., Chen, T., Venugopalan, S., Wang, Z., et al. Is attention all that nerf needs? arXiv preprint arXiv:2207.13298,

    arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.