Pith. sign in

REVIEW 5 major objections 5 minor 60 references

VENI: Variational Encoder for Natural Illumination

T0 review · 5 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read The paper argues that VENI, a rotation-equivariant variational autoencoder built from a Vector Neuron Vision Transformer and a new SO(2)-equivariant layer, gives natural illumination priors a unique, interpolation-friendly latent space that

desk verdict Useful incremental contribution to illumination priors, but the equivariance claim is narrower than advertised and the reconstruction superiority doesn't hold for the autoencoder pass. read the letter →

arxiv 2601.14079 v2 pith:E5PFX7UU submitted 2026-01-20 cs.CV

Paul Walker , James A. D. Gardner , Andreea Ardelean , William A. P. Smith , Bernhard Egger This is my paper
classification cs.CV
keywords rotation-equivariantVAEnaturalilluminationpriorHDRenvironmentmapsVectorNeuronVisionTransformerSO(2)-equivariantlayerlatentspaceuniquenessinterpolationinverserendering
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 sets out to fix a flaw in current natural illumination priors: the decoder-only RENI++ model assigns arbitrary latent codes to similar sky maps, so its latent space is non-unique and interpolation artifacts appear. VENI replaces the autodecoder with a variational autoencoder whose encoder is a Vector Neuron Vision Transformer, built to be equivariant only to rotations around the up-axis, matching the physical symmetry of outdoor lighting. The paper claims this yields a latent space that is unique, reconstruction-consistent, and scalable to large datasets, and that when per-image latent optimization is applied it reproduces or beats RENI++ reconstruction quality. A sympathetic reader would care because a well-behaved illumination prior would make inverse rendering, relighting, and lighting interpolation more reliable and trainable on much more data.

What carries the argument

The key mechanism is the SO(2)-equivariant fully connected layer, an extension of Vector Neurons that treats the x/y components of each direction vector as rotation-equivariant and all other dimensions (z and color) as invariant, combining them bilinearly to preserve equivariance around the up-axis. This layer is used in the Vector Neuron Vision Transformer (VN-ViT) encoder, which processes vertical-stripe patches of the sphere without 2D projection distortion, and whose variational sampling uses an isotropic spherical normal distribution so rotations of the latent code correspond to rotations of the environment map.

What would settle it

As a control, replace VENI's trained encoder with an untrained random projection and recompute the uniqueness metric; if the metric still reports near-zero values, it is not specifically detecting the learned latent-space properties the paper claims. A stronger downstream test: compare recovered lighting accuracy from VENI's forward-pass code versus a per-image optimized RENI++ code on a relighting benchmark; if RENI++ wins without more compute, the claimed practical advantage is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that VENI—a rotation-equivariant variational autoencoder whose encoder is a Vector Neuron Vision Transformer—produces a natural illumination prior with a well-behaved latent space. Unlike the autodecoder prior RENI++, which assigns arbitrary latent codes through per-image optimization, VENI encodes each environment map in a forward pass, so similar images receive similar codes. The authors argue this yields uniqueness (no two codes for the same image), reconstruction consistency (latent distance tracks image distance), smooth interpolation, and scalability to large datasets, while matching or exceeding RENI++ reconstruction quality when latent codes are optimized

Load-bearing premise

The model's central advantage rests on the assumption that its custom uniqueness and reconstruction-consistency metrics measure a latent-space quality that transfers to downstream inverse rendering, rather than being heuristics that overstate the model's practical benefit.

Editorial extensions

If this is right

  • If VENI's latent space is indeed unique, inverse rendering systems can obtain a lighting estimate from a single forward pass and then refine it, rather than optimizing a random code from scratch.
  • Smooth, artifact-free interpolation between environment maps becomes possible, enabling lighting editing and relighting applications.
  • The model scales to much larger datasets than autodecoder approaches, since no per-image latent code optimization is needed during training; this could absorb diverse 360-degree HDR and converted LDR data.
  • The variational architecture offers a distribution over illuminations via the spherical normal prior, which could support sampling-based algorithms like Bayesian inverse rendering or uncertainty-aware relighting.

Reading between the lines

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

  • The paper does not evaluate downstream tasks; a natural next test is whether VENI's encoder-initialized codes speed up inverse rendering convergence and improve final lighting estimates compared with random initialization in RENI++.
  • The custom uniqueness and reconstruction-consistency metrics are heuristic; a stronger validation would compare them against human perceptual judgments of interpolation plausibility or against task-specific accuracy such as recovered albedo or shape error under the predicted lighting.
  • Because the model is trained on a mix of high-quality RENI++ data and HDR-converted street-view LDR images, the synthetic conversion pipeline may bias the prior toward cityscape statistics; extending training to more diverse HDR captures would probe generalization.
  • The SO(2) equivariance choice is principled for outdoor lighting; extending VENI to full SO(3) or to scenes without a canonical up-axis would test whether the equivariance reduction is a genuine advantage or a limitation for other lighting domains.
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

5 major / 5 minor

Summary. The paper proposes VENI, a rotation-equivariant variational autoencoder for natural illumination environment maps. The encoder is a Vector Neuron Vision Transformer (VN-ViT) operating directly on spherical directions, using a novel SO(2)-equivariant fully connected layer and a vertical-stripe patching strategy. The decoder is taken from RENI++. The paper claims that the model is rotation-equivariant, outperforms RENI++ in reconstruction quality, provides a more unique and better-behaved latent space, and scales effectively to large datasets. It evaluates these claims with reconstruction metrics, custom uniqueness/consistency metrics, and a large-scale pretraining study.

Significance. If the central claims held, VENI would be a meaningful advance: an autoencoder illumination prior with an equivariant encoder and unique latent space would overcome RENI++'s autodecoder limitations and enable large-scale training and interpolation. The SO(2)-equivariant fully connected layer and the VN-ViT architecture are interesting, and the equivariance proof in Supplement A is clean. However, several central claims are overstated or contradicted by the paper's own tables, so the current form needs substantial revision.

major comments (5)
  1. [Abstract; §3 Rotation-Equivariant Patching] The paper claims unqualified rotation equivariance in the abstract and conclusion, but the vertical-stripe patching is equivariant only for rotations by multiples of the stripe width (2π/64). The text itself admits this: 'we use 64 patches as a trade-off between memory usage and rotation equivariance.' The proof in Supplement A covers only the FC layer; it does not establish full-model equivariance for arbitrary azimuths. Since equivariance is the method's main differentiator, the claims must be qualified or the patching must be made truly continuous. Please state this limitation in the abstract.
  2. [Table 1; Conclusion] The conclusion states 'Our model outperforms RENI++ in terms of reconstruction quality.' Table 1 shows this is false for the actual autoencoder pass at D=147 and D=300: PSNR 19.40 vs 21.13 and 19.47 vs 22.10; SSIM 0.48 vs 0.51 and 0.48 vs 0.55; LPIPS 0.59 vs 0.55 and 0.59 vs 0.52. Only decoder-only latent optimization beats RENI++, and that variant uses the RENI++ decoder without the proposed encoder. The reconstruction claim should be restricted to the settings where it holds, or the AE gap should be explained.
  3. [Abstract; Table 5] The abstract claims 'our SO(2)-equivariant fully connected layer outperforms standard Vector Neurons when used in our SO(2)-equivariant model.' Table 5 contradicts this for the full SO(2) transformer: at D=147, Full SO(2) gives PSNR 18.18 vs Full VN 18.80; at D=300, 17.93 vs 18.60. The benefit of the new layer appears only when used as projection layers. Please align the claim with the ablation results.
  4. [§4 Uniqueness / Reconstruction Consistency] The central advantage of VENI over RENI++ is the claimed 'well-behaved' latent space. This is supported only by two custom metrics (Table 2) and qualitative interpolations. These heuristics are not validated against any downstream task (e.g., inverse rendering or relighting) or standard distance metrics in the latent space. As the authors state the contribution 'will benefit downstream tasks,' an experiment on at least one downstream task is needed to make the claim load-bearing.
  5. [Table 3] The scaling claim relies on single-run point estimates without variance. At 1500 images, VENI-AE (PSNR 16.00) is substantially worse than RENI++ (20.11); at 43260 it is 16.77 vs RENI++'s 17.14. The trend is in the claimed direction but the absolute performance is still lower, and without error bars the conclusion is fragile. Please report multiple seeds and a baseline trained on the same data for a fair comparison.
minor comments (5)
  1. [Table 1 caption] The caption says 'PSNR, SSIM and LPISP are in...' — 'LPISP' should be 'LPIPS'.
  2. [§3 Rotation-Equivariant Variational Sampling] Eq. (10) uses σ in a way consistent with variance, but the reparameterization text uses σ ambiguously (std vs variance). Please clarify the parameterization.
  3. [Table 5] The table header 'DSO(2) projections' is confusing; it likely means D as a column followed by three model variants. Please format the columns clearly.
  4. [§4 Quality Metrics] The statement 'RENI++ results are taken from their paper, as we found them reproducible' needs detail on how the evaluation protocol (preprocessing, tone-mapping, test split) was matched; otherwise comparisons may be biased.
  5. [General] Architecture and training details (patch size, transformer depth, number of heads, batch size, learning rate, number of epochs) are not fully specified in the main text. Please include them for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity; the central derivation is explicit and benchmarked externally, with a discrete-rotation correctness caveat that is not a circular step.

full rationale

VENI's derivation chain is self-contained. The novel SO(2)-equivariant fully connected layer is defined by explicit equations (Eqs. 1-5), and its equivariance is proven in Supplementary A without assuming the target property. The full encoder's equivariance is limited by the patching scheme: the text explicitly restricts the equivariance result to 'a rotation of the environment map around the up axis by a multiple of the stripe width' (§3), and the statement 'we use 64 patches as a trade-off between memory usage and rotation equivariance' concedes the limitation. That is an overclaim in the abstract relative to the body, but it is a correctness issue, not a circular derivation. The latent-space claims are evaluated empirically against RENI++ using numbers the authors state they reproduced, and the decoder is adopted from RENI++ but benchmarked rather than assumed. The uniqueness and reconstruction-consistency metrics are operational definitions; the paper compares both models on the same procedures instead of defining the conclusion into existence. No equation reduces to a fitted parameter, and no load-bearing premise is justified solely by a self-citation. The self-citations to RENI++/RENI are component reuse with independent quantitative comparison, not circular support.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The paper introduces no new physical entities; the new SO(2)-equivariant fully connected layer and VN-ViT are architectural components, not postulated entities. The main assumptions are domain assumptions about natural illumination (canonical up axis, distant lighting), the fidelity of the LDR-to-HDR conversion, the validity of the scale-invariant training losses, and the meaningfulness of the custom latent-space metrics. A small number of hand-chosen loss weights and architectural constants (patch count, latent dimensions) are free parameters.

free parameters (5)
  • KL divergence weight = 0.01
    Hand-chosen coefficient in Eq. 11; no sensitivity analysis is reported.
  • MAGE loss weight = 0.5
    Hand-chosen coefficient in Eq. 11; no sensitivity analysis is reported.
  • Scale-invariant and cosine loss weights = 1.0 each
    Hand-chosen coefficients in Eq. 11; no sensitivity analysis is reported.
  • Number of vertical stripe patches = 64
    Chosen as a memory/equivariance trade-off in Section 3; no ablation varies patch count.
  • Latent dimension D = 27, 147, 300
    Evaluation points; no explicit selection criterion or hyperparameter search is given for the final model.
assumptions (7)
  • domain assumption Natural outdoor illumination environments have a canonical up-axis, and only rotations around it map one realistic environment to another.
    Section 3 motivates SO(2) rather than SO(3) equivariance; if scenes without a clear horizon/canonical up are included, the equivariance group choice is wrong.
  • domain assumption Environment maps are valid distant-illumination spherical signals shared across a scene.
    Standard computer graphics assumption invoked by the choice of environment-map representation in Section 2.
  • domain assumption The LDR-to-HDR conversion method of [51] produces training data of sufficient radiometric fidelity for pretraining.
    Section 3 uses 43,310 streetlearn images converted to HDR; the paper acknowledges the conversion is imperfect.
  • domain assumption Scale-invariant losses and log-space training correctly resolve unknown exposure/scale ambiguity.
    Section 4 relies on log-space training and scale-invariant losses to factor out unknown EV; if scale ambiguity remains, the prior may capture spurious scale variation.
  • standard math VN-Transformer and the RENI++ decoder are valid, and their equivariance properties compose with the new SO(2) layer.
    The proof in Supplement A covers the new fully connected layer; the full-stack equivariance depends on the cited equivariant attention and decoder.
  • ad hoc to paper The fixed vertical-stripe patching with per-pixel direction vectors is SO(2)-equivariant for arbitrary rotations, not only rotations by multiples of the stripe width.
    Section 3 'Rotation-Equivariant Patching' explicitly proves only the stripe-width-multiple case in the text, and relies on the presence of direction vectors to extend equivariance; this extension is plausible but not stated as a formal theorem.
  • ad hoc to paper The custom uniqueness and reconstruction-consistency metrics measure latent-space well-behavedness relevant to downstream tasks.
    Section 4 introduces these metrics as the main evidence for 'well-behaved' latent space; no formal or downstream-task validation is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VENI: Variational Encoder for Natural Illumination." pith.science (2026). https://pith.science/paper/E5PFX7UU

@misc{pith2026260114079,
  author       = {Pith},
  title        = {Pith review of: VENI: Variational Encoder for Natural Illumination},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E5PFX7UU}},
  note         = {Machine review of arXiv:2601.14079}
}
read the original abstract

Inverse rendering is an ill-posed problem, but priors such as illumination priors can help simplify it. Existing work either disregards the spherical and rotation-equivariant nature of illumination environments or does not provide a well-behaved latent space. We propose a rotation-equivariant variational autoencoder that models natural illumination on the sphere without relying on 2D projections. To preserve the SO(2)-equivariance of environment maps, we use a novel Vector Neuron Vision Transformer (VN-ViT) as encoder and a rotation-equivariant conditional neural field as decoder. In the encoder, we reduce the equivariance from SO(3) to SO(2) using a novel SO(2)-equivariant fully connected layer, an extension of Vector Neurons. We show that our SO(2)-equivariant fully connected layer outperforms standard Vector Neurons when used in our SO(2)-equivariant model. Compared to previous methods, our variational autoencoder enables smoother interpolation in latent space and offers a more well-behaved latent space.

Figures

Figures reproduced from arXiv: 2601.14079 by the authors.

Figure 1
Figure 1. We build a rotation-equivariant variational autoencoder model to address the limitations of the current state of-the-art illumination [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Our model is rotation-equivariant, meaning that a rotation [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Model Overview: We adapt the Vision Transformer architecture to SO(2) equivariance on spherical images. Splitting the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Vertical stripes as patches is a valid SO(2)-equivariant [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Interpolations using our model (with direct latent opti [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Interpolating between two optimized latent codes. With our model, the output images remain consistent during interpolation, [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Interpolations using our model (with direct latent optimization) and RENI++ with different latent sizes. Three image pairs are [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Interpolations using our model (autoencoder pass) and RENI++ with different latent sizes. Three image pairs are interpolated: [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Corresponding figure to Table [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Corresponding figure to Table [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 12
Figure 12. Figure 12: Corresponding figure to Table [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: More examples following Figure [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 2 linked inside Pith

  1. [1]

    Vn-transformer: Rotation- equivariant attention for vector neurons.Trans

    Serge Assaad, Carlton Downey, Rami Al-Rfou’, Nigamaa Nayakanti, and Benjamin Sapp. Vn-transformer: Rotation- equivariant attention for vector neurons.Trans. Mach. Learn. Res., 2023, 2023. 3, 4, 1

    2023

  2. [2]

    Barron and Jitendra Malik

    Jonathan T. Barron and Jitendra Malik. Intrinsic scene prop- erties from a single rgb-d image. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013. 1, 2

    2013

  3. [3]

    Barron and Jitendra Malik

    Jonathan T. Barron and Jitendra Malik. Shape, illumination, and reflectance from shading.IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(8):1670–1687, 2015. 1, 2, 3

    2015

  4. [4]

    Basri and D.W

    R. Basri and D.W. Jacobs. Lambertian reflectance and lin- ear subspaces.IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(2):218–233, 2003. 1, 2

    2003

  5. [5]

    The bas-relief ambiguity.International journal of computer vision, 35(1):33–44, 1999

    Peter N Belhumeur, David J Kriegman, and Alan L Yuille. The bas-relief ambiguity.International journal of computer vision, 35(1):33–44, 1999. 1

    1999

  6. [6]

    Richter, and Vladlen Koltun

    Aleksei Bochkovskii, Ama¨el Delaunoy, Hugo Germain, Mar- cel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second, 2025. 6

    2025

  7. [7]

    Neural-pil: Neural pre- integrated lighting for reflectance decomposition

    Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik PA Lensch. Neural-pil: Neural pre- integrated lighting for reflectance decomposition. InAd- vances in Neural Information Processing Systems, pages 10691–10704. Curran Associates, Inc., 2021. 1, 2

    2021

  8. [8]

    Bronstein, Joan Bruna, Taco Cohen, and Petar Velickovic

    Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Velickovic. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges.CoRR, abs/2104.13478, 2021. 3

    arXiv 2021

Show all 60 references
  1. [9]

    Cohen and Max Welling

    Taco S. Cohen and Max Welling. Steerable CNNs. InIn- ternational Conference on Learning Representations, 2017. 3

    2017

  2. [10]

    Ev- erlight: Indoor-outdoor editable hdr lighting estimation

    Mohammad Reza Karimi Dastjerdi, Jonathan Eisenmann, Yannick Hold-Geoffroy, and Jean-Fran c ¸ois Lalonde. Ev- erlight: Indoor-outdoor editable hdr lighting estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7420–7429, 2023. 2, 3

    2023

  3. [11]

    Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J. Guibas. Vector neu- rons: A general framework for so(3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12200–12209, 2021....

    2021

  4. [12]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

    2021

  5. [13]

    Dror, Alan S

    Ron O. Dror, Alan S. Willsky, and Edward H. Adelson. Sta- tistical characterization of real-world illumination.Journal of Vision, 4(9):11–11, 2004. 1, 2

    2004

  6. [14]

    Occlusion-aware 3d morphable models and an illumination prior for face image analysis.Interna- tional Journal of Computer Vision, 126(12):1269–1287, 2018

    Bernhard Egger, Sandro Sch ¨onborn, Andreas Schneider, Adam Kortylewski, Andreas Morel-Forster, Clemens Blumer, and Thomas Vetter. Occlusion-aware 3d morphable models and an illumination prior for face image analysis.Interna- tional Journal of Computer Vision, 126(12):1269–128...

    2018

  7. [15]

    Depth map prediction from a single image using a multi-scale deep network, 2014

    David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network, 2014. 6

    2014

  8. [16]

    Mantiuk, and Jonas Unger

    Gabriel Eilertsen, Joel Kronander, Gyorgy Denes, Rafał K. Mantiuk, and Jonas Unger. Hdr image reconstruction from a single exposure using deep cnns.ACM Trans. Graph., 36(6),

  9. [17]

    Real-world illumination and the perception of surface re- flectance properties.Journal of vision, 3(5):3–3, 2003

    Roland W Fleming, Ron O Dror, and Edward H Adelson. Real-world illumination and the perception of surface re- flectance properties.Journal of vision, 3(5):3–3, 2003. 1

    2003

  10. [18]

    Se(3)-transformers: 3d roto-translation equivari- ant attention networks

    Fabian Fuchs, Daniel Worrall, V olker Fischer, and Max Welling. Se(3)-transformers: 3d roto-translation equivari- ant attention networks. InAdvances in Neural Information Processing Systems, pages 1970–1981. Curran Associates, Inc., 2020. 3

    1970

  11. [19]

    Rotation- equivariant conditional spherical neural fields for learning a natural illumination prior

    James Gardner, Bernhard Egger, and William Smith. Rotation- equivariant conditional spherical neural fields for learning a natural illumination prior. InAdvances in Neural Information Processing Systems, pages 26309–26323. Curran Associates, Inc., 2022. 1, 2, 3

    2022

  12. [20]

    James A. D. Gardner, Bernhard Egger, and William A. P. Smith. Reni++ a rotation-equivariant, scale-invariant, natural illumination prior, 2023. 1, 2, 3, 5, 6

    2023

  13. [21]

    James A. D. Gardner, Evgenii Kashin, Bernhard Egger, and William A. P. Smith. The sky’s the limit: Relightable out- door scenes via a sky-pixel constrained illumination prior and outside-in visibility. InComputer Vision – ECCV 2024, pages 126–143, Cham, 2025. Springer Nature S...

    2024

  14. [22]

    Learning to predict indoor illumination from a single image.ACM Trans

    Marc-Andr´e Gardner, Kalyan Sunkavalli, Ersin Yumer, Xiao- hui Shen, Emiliano Gambaretto, Christian Gagn´e, and Jean- Franc ¸ois Lalonde. Learning to predict indoor illumination from a single image.ACM Trans. Graph., 36(6), 2017. 2

    2017

  15. [23]

    Environment mapping and other applications of world projections.IEEE Computer Graphics and Applica- tions, 6(11):21–29, 1986

    Ned Greene. Environment mapping and other applications of world projections.IEEE Computer Graphics and Applica- tions, 6(11):21–29, 1986. 2

    1986

  16. [24]

    Rehg, and Varun Jampani

    Zixuan Huang, Mark Boss, Aaryaman Vasishta, James M. Rehg, and Varun Jampani. Spar3d: Stable point-aware recon- struction of 3d objects from single images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16860–16870, 2025. 1

    2025

  17. [25]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401–4410, 2019. 2 9

    2019

  18. [26]

    Auto-encoding varia- tional bayes

    Diederik P Kingma and Max Welling. Auto-encoding varia- tional bayes. InInternational Converence on Learnin Repre- sentations (ICLR), 2014. 4

    2014

  19. [27]

    Unveiling the ambiguity in neural inverse rendering: A parameter compensation analysis

    Georgios Kouros, Minye Wu, Sushruth Nagesh, Xianling Zhang, and Tinne Tuytelaars. Unveiling the ambiguity in neural inverse rendering: A parameter compensation analysis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages...

    2024

  20. [28]

    Megadepth: Learning single- view depth prediction from internet photos

    Zhengqi Li and Noah Snavely. Megadepth: Learning single- view depth prediction from internet photos. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6

    2018

  21. [29]

    The streetlearn environment and dataset, 2019

    Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, Denis Teplyashin, Karl Moritz Hermann, Mateusz Mali- nowski, Matthew Koichi Grimes, Karen Simonyan, Koray Kavukcuoglu, Andrew Zisserman, and Raia Hadsell. The streetlearn environment and dataset, 2019. 2, 5

    2019

  22. [30]

    Visual perception and natural illumination.Current Opinion in Behavioral Sciences, 30:48–54, 2019

    Richard F Murray and Wendy J Adams. Visual perception and natural illumination.Current Opinion in Behavioral Sciences, 30:48–54, 2019. Visual perception. 1

    2019

  23. [31]

    Deepsdf: Learning continuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard New- combe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165–174, 2019. 2

    2019

  24. [32]

    Diffusionlight: Light probes for free by painting a chrome ball

    Pakkapon Phongthawee, Worameth Chinchuthakun, Non- taphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pat...

    2024

  25. [33]

    A functional approach to rotation equivariant non-linearities for tensor field networks

    Adrien Poulenard and Leonidas J Guibas. A functional approach to rotation equivariant non-linearities for tensor field networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13174– 13183, 2021. 3

    2021

  26. [34]

    An efficient repre- sentation for irradiance environment maps

    Ravi Ramamoorthi and Pat Hanrahan. An efficient repre- sentation for irradiance environment maps. InProceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, page 497–500, New York, NY , USA,

  27. [35]

    Matthews, Kwang Moo Yi, Gopal Sharma, Dmitry Lagun, and Andrea Tagliasacchi

    Daniel Rebain, Mark J. Matthews, Kwang Moo Yi, Gopal Sharma, Dmitry Lagun, and Andrea Tagliasacchi. Attention beats concatenation for conditioning neural fields, 2022. 5

    2022

  28. [36]

    Numerische isotropieoptimierung von fir-filtern mittels quergl¨attung

    Hanno Scharr, Stefan K¨orkel, and Bernd J¨ahne. Numerische isotropieoptimierung von fir-filtern mittels quergl¨attung. In Mustererkennung 1997, pages 367–374, Berlin, Heidelberg,

    1997

  29. [37]

    Panoformer: Panorama transformer for indoor 360° depth estimation

    Zhijie Shen, Chunyu Lin, Kang Liao, Lang Nie, Zishuo Zheng, and Yao Zhao. Panoformer: Panorama transformer for indoor 360° depth estimation. InComputer Vision – ECCV 2022, pages 195–211, Cham, 2022. Springer Nature Switzer- land. 3

    2022

  30. [38]

    Hdr map reconstruction from a single ldr sky panoramic image for outdoor illumination estimation

    Gyeongik Shin, Kyeongmin Yu, Mpabulungi Mark, and Hyunki Hong. Hdr map reconstruction from a single ldr sky panoramic image for outdoor illumination estimation. IEEE Access, 11:17359–17374, 2023. 5, 6

    2023

  31. [39]

    Stone, I.S

    J.V . Stone, I.S. Kerrigan, and J. Porrill. Where is the light? bayesian perceptual priors for lighting direction.Proceedings of the Royal Society B: Biological Sciences, 276(1663):1797– 1804, 2009. 1

    2009

  32. [40]

    High-dynamic-range lighting estima- tion from face portraits

    Alejandro Sztrajman, Alexandros Neophytou, Tim Weyrich, and Eric Sommerlade. High-dynamic-range lighting estima- tion from face portraits. In2020 International Conference on 3D Vision (3DV), pages 355–363, 2020. 2, 3

    2020

  33. [41]

    Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds, 2018

    Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds, 2018. 3

    2018

  34. [42]

    light-from-above

    Rhiannon Thomas, Marko Nardini, and Denis Mareschal. Interactions between “light-from-above” and convexity priors in visual development.Journal of Vision, 10(8):6–6, 2010. 1

    2010

  35. [43]

    Dist-gan: An improved gan using distance constraints

    Ngoc-Trung Tran, Tuan-Anh Bui, and Ngai-Man Cheung. Dist-gan: An improved gan using distance constraints. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 8

    2018

  36. [44]

    All-frequency precom- puted radiance transfer using spherical radial basis functions and clustered tensor approximation.ACM Trans

    Yu-Ting Tsai and Zen-Chung Shih. All-frequency precom- puted radiance transfer using spherical radial basis functions and clustered tensor approximation.ACM Trans. Graph., 25 (3):967–976, 2006. 2

    2006

  37. [45]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkor- eit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neu- ral Information Processing Systems. Curran Associates, Inc.,

  38. [46]

    Lediff: Latent exposure diffusion for hdr generation

    Chao Wang, Zhihao Xia, Thomas Leimkuhler, Karol Myszkowski, and Xuaner Zhang. Lediff: Latent exposure diffusion for hdr generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 453–464, 2025. 5, 6

    2025

  39. [47]

    Stylelight: Hdr panorama generation for lighting estima- tion and editing

    Guangcong Wang, Yinuo Yang, Chen Change Loy, and Ziwei Liu. Stylelight: Hdr panorama generation for lighting estima- tion and editing. InComputer Vision – ECCV 2022, pages 477–492, Cham, 2022. Springer Nature Switzerland. 2, 3

    2022

  40. [48]

    Materialist: Physically based editing using single-image inverse rendering.arXiv preprint arXiv:2501.03717, 2025

    Lezhong Wang, Duc Minh Tran, Ruiqi Cui, Thomson TG, An- ders Bjorholm Dahl, Siavash Arjomand Bigdeli, Jeppe Revall Frisvad, and Manmohan Chandraker. Materialist: Physically based editing using single-image inverse rendering.arXiv preprint arXiv:2501.03717, 2025. 1, 2

    2025 arXiv

  41. [49]

    Learn- ing indoor inverse rendering with 3d spatially-varying light- ing

    Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learn- ing indoor inverse rendering with 3d spatially-varying light- ing. InProceedings of the IEEE/CVF International Confer- ence on Computer Vision (ICCV), pages 12538–12547, 2021. 1, 2

    2021

  42. [50]

    3d steerable cnns: Learn- ing rotationally equivariant features in volumetric data

    Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco S Cohen. 3d steerable cnns: Learn- ing rotationally equivariant features in volumetric data. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3

    2018

  43. [51]

    Luminance attentive networks for hdr image and panorama reconstruction.Computer Graphics Forum, 40(7):181–192, 2021

    Hanning Yu, Wentao Liu, Chengjiang Long, Bo Dong, Qin Zou, and Chunxia Xiao. Luminance attentive networks for hdr image and panorama reconstruction.Computer Graphics Forum, 40(7):181–192, 2021. 5, 6

    2021

  44. [52]

    Ye Yu and William A. P. Smith. Outdoor inverse rendering from a single image using multiview self-supervision.IEEE 10 Transactions on Pattern Analysis and Machine Intelligence, 44(7):3659–3675, 2022. 1, 2, 3

    2022

  45. [53]

    Laformer: Vision transformer for panoramic image semantic segmentation.IEEE Signal Processing Letters, 30: 1792–1796, 2023

    Zheng Yuan, Junhua Wang, Yuxin Lv, Ding Wang, and Yi Fang. Laformer: Vision transformer for panoramic image semantic segmentation.IEEE Signal Processing Letters, 30: 1792–1796, 2023. 3

    2023

  46. [54]

    Panoramic vision transformer for saliency detection in 360° videos

    Heeseung Yun, Sehun Lee, and Gunhee Kim. Panoramic vision transformer for saliency detection in 360° videos. In Computer Vision – ECCV 2022, pages 422–439, Cham, 2022. Springer Nature Switzerland. 3

    2022

  47. [55]

    Emlight: Lighting estimation via spherical distribution approximation

    Fangneng Zhan, Changgong Zhang, Yingchen Yu, Yuan Chang, Shijian Lu, Feiying Ma, and Xuansong Xie. Emlight: Lighting estimation via spherical distribution approximation. Proceedings of the AAAI Conference on Artificial Intelligence, 35(4):3287–3295, 2021. 2, 3

    2021

  48. [56]

    Physg: Inverse rendering with spherical gaus- sians for physics-based material editing and relighting

    Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaus- sians for physics-based material editing and relighting. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5453–...

    2021

  49. [57]

    Modeling indirect illumination for in- verse rendering

    Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for in- verse rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18643–18652, 2022. 1, 2

    2022

  50. [58]

    Distortion-aware transformer in 360° salient object detection

    Yinjie Zhao, Lichen Zhao, Qian Yu, Lu Sheng, Jing Zhang, and Dong Xu. Distortion-aware transformer in 360° salient object detection. InProceedings of the 31st ACM Interna- tional Conference on Multimedia, page 499–508, New York, NY , USA, 2023. Association for Computing Machin...

    2023

  51. [1997]

    Springer Berlin Heidelberg. 6

  52. [2001]

    Association for Computing Machinery. 2

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.