Pith. sign in

REVIEW 2 major objections 4 minor 123 references

Global illumination in volumetric inverse rendering can be enforced without ever running a global-illumination renderer: a pointwise radiative-transfer residual plus a primary-ray volume-rendering term suffice to recover physically meaningf

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 03:59 UTC pith:Y5Q55WGL

arxiv 2607.13695 v1 pith:Y5Q55WGL submitted 2026-07-15 cs.GR

Volumetric Inverse Rendering via Neural Radiative Transfer

Ntumba Elie Nsampi , Adarsh Djeacoumar , Hans-Peter Seidel , Tobias Ritschel , Thomas Leimk\"uhler This is my paper
classification cs.GR
keywords volumetric inverse renderingradiative transfer equationglobal illuminationneural fieldsparticipating mediaphase functionphysics-informed optimizationgenerative model
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that recovering a scattering, absorbing volume's optical properties under full global illumination does not require differentiating a Monte Carlo path tracer. Instead, two neural fields — one for the medium (color-resolved absorption, scattering, and a Henyey–Greenstein phase parameter) and one for the full spatio-angular light field — are optimized together against a pointwise residual of the Radiative Transfer Equation, together with boundary conditions, multi-view image observations, and a volume-rendering integral along primary rays. The local RTE residual is the engine that implicitly accounts for multiple scattering; the primary-ray VRE term counters the low-frequency smoothness that pure differential constraints produce. The paper demonstrates this on synthetic scenes with known illumination, recovering parameters that support relighting and novel views, and extends the same machinery to a generative model of participating media.

Core claim

The central claim is that a global-illumination renderer is not essential for global-illumination-aware inverse rendering. Representing the medium's optical properties and the scene's radiance field as neural fields and constraining them with the local, pointwise radiative-transfer residual (plus inflow boundary conditions, camera observation constraints, and a volume-rendering term along viewing rays) drives the joint optimization to a globally consistent light-transport equilibrium. The in-scattering integral is evaluated by Monte Carlo sampling of directions at collocation points, so no ray tracing through the volume is needed; the VRE term provides nonlocal coupling that restores high-fr

What carries the argument

The machinery is the local differential Radiative Transfer Equation written as a residual objective: at randomly sampled collocation points (x, ω), the directional derivative of the neural light field must balance extinction, emission, and Monte-Carlo-evaluated in-scattering from the neural medium field. A boundary-condition term fixes the inflow radiance from known environment illumination, and an observation term pins the light field at camera positions. The complementary Volume Rendering Equation term integrates radiance along primary viewing rays only, using the jointly optimized light field for in-scattering; this is the component that prevents low-frequency bias and anchors high-freque

Load-bearing premise

The load-bearing premise is that a pointwise RTE residual, evaluated at 10,000 collocation points with 32 Monte Carlo directions, plus sparse camera observations and known boundary inflow, is enough to force the jointly optimized light field and medium to converge to the true optical properties, with no convergence or uniqueness guarantee.

What would settle it

Render a synthetic volume with known absorption, scattering, and Henyey–Greenstein parameter, run the method, and measure the recovered parameter error while increasing the number of camera views and collocation samples. If the error does not decrease toward zero, or if a second medium that produces the same multi-view images under the same illumination is found and the method recovers the wrong one with a near-zero RTE residual, then the joint optimization is not sufficient to identify the true medium.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Inverse rendering of participating media can be implemented with general-purpose neural optimization, without custom differentiable path-tracing code or variance-reduction schemes.
  • The recovered parameters are physical, so the reconstructed volume can be re-rendered under novel illumination and from novel views, including scenes with anisotropic phase functions.
  • The same joint neural-field formulation extends to a generative model: a learned distribution over latent codes yields new volumes whose optical properties are physically consistent under global illumination.
  • Because the formulation is compositional, it can be extended to spectral and transient light transport, dynamic scenes, and light sources modeled by an emission term.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The formulation invites a direct test of identifiability: sweep optical depth and phase-function anisotropy while holding the image count fixed and measure whether recovery error stays bounded; if error grows sharply, the local RTE residual is not disambiguating as strongly as the examples suggest.
  • The generative auto-decoder learns a distribution over medium–light-field pairs; we infer that the same machinery could serve as a learned prior for inverse rendering from a single photograph, initializing the medium field and a latent code rather than starting from scratch.
  • Since the VRE term dominates runtime, a cheap extension implied by the paper's own analysis is to sample primary rays adaptively or to add a small number of off-axis rays, which would test whether the high-frequency anchoring is specific to primary viewing directions.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes a volumetric inverse rendering method that represents both the medium's optical properties (absorption σa, scattering σs, and a spatially varying Henyey-Greenstein phase parameter g) and the scene's full radiance field as neural fields. These are jointly optimized under a physics-informed objective: an RTE residual at collocation points, inflow boundary conditions, pixel observations, and a primary-ray volume-rendering data term. The authors demonstrate reconstruction on 50 synthetic multi-view scenes with ground truth, compare against Differential Ratio Tracking and TensorIR, ablate the VRE term, and extend the framework to a generative auto-decoder model over volumetric scenes. The central claim is that global illumination can be accounted for in inverse rendering without running a global-illumination renderer.

Significance. If the method is correct, it is a significant step: it avoids the engineering and variance burden of differentiable Monte Carlo light transport while still recovering physically interpretable optical properties under multiple scattering. The paper is honest about its synthetic-only evaluation and the lack of real datasets. It ships code and models, uses a standard physics model (RTE), and provides a useful baseline comparison. The main risk is a concrete technical error in the primary-ray data term (Eq. 7, inherited from Eq. 2), which is load-bearing because that term has the largest weight in the objective. The empirical results also lack error bars, which weakens the quantitative claims.

major comments (2)
  1. [Sec. 3, Eq. (2); Sec. 4.1, Eq. (7); Sec. 4.3] The integral form of the RTE in Eq. (2) is incomplete. For a finite volume with nonzero inflow, the standard VRE must include the attenuated boundary radiance T(0,t_b)L(x_b,ω). Eq. (7) inherits this omission: the primary-ray data term compares the observed pixel I_k(q) with only the in-scattering integral along the ray, omitting T(0,t_b)L_in(x_b,-ω). Since the scenes are illuminated by environment maps and include weakly absorbing, semi-transparent elements (Sec. 5), the transmitted background is generally nonzero. With λ_VRE=70, minimizing Eq. (7) forces the in-scattering integral alone to reproduce the observed radiance, which can bias σs and the phase function upward or distort g to fabricate radiance that actually comes from the far boundary. This changes the optimization target and could invalidate the reported parameter MSEs. Please correct Eq. (7) (and Eq. (2)), state whether the
  2. [Sec. 5.1, Tables 1 and 2] All quantitative claims rest on a single aggregate MSE/PSNR/SSIM/FID value per method and condition. The text describes the σs/σt advantage over DRT as 'significantly more accurate' (Sec. 5.1), but with 50 scenes and no per-scene variance, standard deviation, or confidence intervals, the reader cannot assess whether this is systematic or driven by outliers. Please report per-scene statistics (mean±std, box plots, or bootstrap intervals) and, if feasible, multiple optimization runs with different seeds. This is directly relevant to the paper's central empirical claim.
minor comments (4)
  1. [Sec. 5.1] The weighting of the g error by σs is described only in prose. Please define whether σs is aggregated over RGB channels and how the weighting behaves when σs≈0.
  2. [Sec. 4.3] The value λ_VRE=70 is a single hand-picked hyperparameter. A small sensitivity analysis (e.g., λ_VRE ∈ {10, 30, 70, 150}) would help the reader understand how robust the method is to this balance.
  3. [Sec. 5] The synthetic dataset spans a wide range of optical depths, but the paper does not report statistics of the boundary transmittance in the test views. Reporting these would directly address the magnitude of the omitted boundary term and help interpret the corrected experiments.
  4. [Sec. 4.1] The paper does not discuss identifiability: the recovery of σa, σs, and g separately from images is an inverse problem with potential ambiguities. A simple analytical test case (e.g., a homogeneous slab with known environment illumination) would strengthen the claim that the three parameters are separately recoverable.

Circularity Check

0 steps flagged

No significant circularity: the reconstruction is an inverse optimization constrained by standard physics, and the reported predictions are evaluated against independently path-traced ground truth.

full rationale

The derivation chain is self-contained. Eq. 1 is the standard radiative transfer equation, and Eq. 3 is a direct pointwise residual of that physical law; boundary and observation terms (Eqs. 5–7) incorporate independent data, namely environment illumination and multi-view images. The medium properties (σa, σs, g) and light field Lθ are jointly optimized by minimizing Eq. 8; no parameter is fitted to the held-out test views or to the relighting target. The quantitative evaluation in Sec. 5 uses ground-truth optical properties from synthetic scenes and synthesizes novel-view/relighting images by re-rendering the reconstructed volumes with volumetric path tracing, so the reported predictions are not equal to the optimization inputs by construction. The fact that synthetic training data were generated with the same RTE is a standard self-consistency check, not circularity. The possible omission of a boundary/transmission term in the VRE objective (Eq. 7) and the absence of a theoretical uniqueness/convergence proof are correctness and well-posedness concerns, not circular-reasoning defects. Self-citations appearing in related work are descriptive and are not used to justify the central optimization; no load-bearing claim rests on a self-citation chain.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical entities or forces. It relies on standard radiative transfer physics and makes pragmatic modeling choices (single-parameter HG phase function, no emission, known environment lighting). The free parameters are hand-chosen loss weights and a latent dimension.

free parameters (2)
  • lambda_VRE = 70
    Loss weight for the volume rendering objective (Eq. 7), chosen by hand to balance the terms in Eq. 8; ablation only tests its presence/absence, not sensitivity.
  • latent_dim dz = 4
    Dimensionality of the generative latent code in Sec. 4.2, chosen for the proof-of-concept; no analysis of its effect.
axioms (5)
  • standard math The steady-state RTE (Eq. 1) correctly models light transport in the scenes.
    Standard physics, used throughout; evaluated data is rendered with the same equation, so synthetic validation is self-consistent.
  • domain assumption The light field is well-approximated by a neural field L_theta with sufficient capacity to represent the 5D radiance distribution.
    Necessary for the RTE residual and VRE integrals to be accurate; no capacity bound or error analysis is given (Sec. 4.1).
  • ad hoc to paper Henyey-Greenstein phase function with a single spatially varying parameter g suffices for the scattering media.
    Simplifies the scattering model; real media may require more complex phase functions (Sec. 4.1).
  • domain assumption The inverse problem is well-posed: the RTE residual plus boundary and observation constraints are sufficient to disambiguate absorption, scattering, and phase function.
    The paper does not prove uniqueness or stability; it relies on the empirical success on synthetic scenes (Sec. 4.1, Sec. 5).
  • domain assumption Environment illumination is known and fixed; no emission inside the volume.
    The method omits the emission term in Eq. 3; the paper states this is an extension for future work (Sec. 4.1, Sec. 6).

pith-pipeline@v1.3.0-alltime-deepseek · 16807 in / 10970 out tokens · 104389 ms · 2026-08-02T03:59:34.257869+00:00 · methodology

0 comments
read the original abstract

Volumetric inverse rendering seeks to recover the optical properties of participating media from images. Existing approaches either rely on differentiable stochastic light transport simulation, which require substantial algorithmic effort, or use simplified models that fail to capture global illumination. We propose a formulation that reconciles physically complete light transport with general-purpose neural optimization. The optical properties of the medium and the full light field are represented as neural fields and estimated through a joint optimization process. Global illumination is enforced via a residual objective derived from the Radiative Transfer Equation in local differential form, complemented by a volume rendering term along primary viewing rays to mitigate \rev{low-frequency} bias. We demonstrate reconstruction of spatially varying, color-resolved scattering, absorption, and phase function parameters from multi-view images. Beyond reconstruction, the same framework supports learning generative models of participating media with physical optical properties under global illumination.

Figures

Figures reproduced from arXiv: 2607.13695 by Adarsh Djeacoumar, Hans-Peter Seidel, Ntumba Elie Nsampi, Thomas Leimk\"uhler, Tobias Ritschel.

Figure 1
Figure 1. Figure 1: We propose a formulation for volumetric inverse rendering under global illumination without explicit global-illumination rendering. By framing the problem as constrained optimization over neural fields, the approach enables physically consistent reconstruction, novel-view synthesis, and relighting. Here, we demonstrate relighting of the reconstructed volume with anisotropic optical properties using a combi… view at source ↗
Figure 2
Figure 2. Figure 2: Different approaches to volumetric inverse rendering. (a) The emission–absorption model bakes illumination into the vol￾ume, preventing recovery of physically meaningful optical prop￾erties. (b) Restricting light transport to a single bounce enables limited relighting but fails to capture the complexity of global illu￾mination. (c) Differentiable stochastic path sampling accounts for global illumination du… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our approach. Two neural fields encode the optical properties of the participating medium and the scene’s light field (orange). Disentanglement is guided by multiple optimization objectives (grey) that incorporate the available data (green). Here, points along the ray x+τω are denoted by xτ for τ ≥ 0, and tb is the distance to the volume boundary. The volume rendering equation expresses radianc… view at source ↗
Figure 4
Figure 4. Figure 4: Reconstruction results on two scenes (row blocks), comparing different methods (rows). The first three columns show linearly tonemapped slices through the reconstructed medium properties, namely absorption (σa), scattering (σs), and extinction (σt). The remaining columns demonstrate novel-view synthesis under novel illumination. TensorIR does not recover volumetric medium properties and is therefore shown … view at source ↗
Figure 5
Figure 5. Figure 5: Three scenes (columns) sampled from our generative model, showing diverse volumetric structures with physically mean￾ingful optical properties, rendered under training (top row) and novel (bottom row) illumination. All images are rendered from the same view. w/o VRE w/ VRE Ground Truth [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Reconstruction of scattering parameters σs (shown here as representative of all recovered volumetric parameters) on a slice through the volume. Omitting the VRE objective (Eq. 7) leads to overly smooth reconstructions, whereas incorporating it yields re￾sults closer to the ground truth. to other lighting representations, such as local light sources, by in￾corporating the emission term into the radiative tr… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

123 extracted references · 3 linked inside Pith

  1. [1]

    ACM Transactions on Graphics (ToG) , volume=

    Physics informed neural fields for smoke reconstruction with sparse data , author=. ACM Transactions on Graphics (ToG) , volume=. 2022 , publisher=

    2022

  2. [2]

    Advances in Neural Information Processing Systems , volume=

    Inferring hybrid neural fluid fields from videos , author=. Advances in Neural Information Processing Systems , volume=

  3. [3]

    Journal of Computational physics , volume=

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , author=. Journal of Computational physics , volume=. 2019 , publisher=

    2019

  4. [4]

    ACM Transactions on Graphics (TOG) , volume=

    Inverse volume rendering with material dictionaries , author=. ACM Transactions on Graphics (TOG) , volume=. 2013 , publisher=

    2013

  5. [5]

    ACM Transactions on Graphics (TOG) , volume=

    Scalarflow: a large-scale volumetric data set of real-world scalar transport flows for computer animation and machine learning , author=. ACM Transactions on Graphics (TOG) , volume=. 2019 , publisher=

    2019

  6. [6]

    Computer graphics forum , volume=

    Transparent and specular object reconstruction , author=. Computer graphics forum , volume=. 2010 , organization=

    2010

  7. [7]

    Advances in Neural Information Processing Systems , volume=

    Neural relightable participating media rendering , author=. Advances in Neural Information Processing Systems , volume=

  8. [8]

    The Thirteenth International Conference on Learning Representations , year=

    Implicit Neural Surface Deformation with Explicit Velocity Fields , author=. The Thirteenth International Conference on Learning Representations , year=

  9. [9]

    Computer graphics forum , volume=

    Survey of models for acquiring the optical properties of translucent materials , author=. Computer graphics forum , volume=. 2020 , organization=

    2020

  10. [10]

    arXiv preprint arXiv:2303.05512 , year=

    Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification , author=. arXiv preprint arXiv:2303.05512 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:1905.11169 , year=

    Physics-as-inverse-graphics: Unsupervised physical parameter estimation from video , author=. arXiv preprint arXiv:1905.11169 , year=

    Pith/arXiv arXiv 1905

  12. [12]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Tomofluid: Reconstructing dynamic fluid from sparse view videos , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  13. [13]

    ACM SIGGRAPH 2006 Papers , pages=

    Acquiring scattering properties of participating media by dilution , author=. ACM SIGGRAPH 2006 Papers , pages=

    2006

  14. [14]

    European Conference on Computer Vision , pages=

    Regularizing Dynamic Radiance Fields with Kinematic Fields , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  15. [15]

    Advances in Neural Information Processing Systems , volume=

    NVFi: neural velocity fields for 3D physics learning from dynamic videos , author=. Advances in Neural Information Processing Systems , volume=

  16. [16]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , year=

    Nerv: Neural reflectance and visibility fields for relighting and view synthesis , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , year=

  17. [17]

    arXiv preprint arXiv:2008.03824 , year=

    Neural reflectance fields for appearance acquisition , author=. arXiv preprint arXiv:2008.03824 , year=

    Pith/arXiv arXiv 2008

  18. [18]

    ACM Transactions on Graphics (TOG) , volume=

    NeuralTO: Neural Reconstruction and View Synthesis of Translucent Objects , author=. ACM Transactions on Graphics (TOG) , volume=. 2024 , publisher=

    2024

  19. [19]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages=

    FluidGS: Physics Informed Gaussian Splatting for Dynamic Fluid Reconstruction from Sparse Views , author=. Proceedings of the 33rd ACM International Conference on Multimedia , pages=

  20. [20]

    ACM Transactions on Graphics (ToG) , volume=

    Nerfactor: Neural factorization of shape and reflectance under an unknown illumination , author=. ACM Transactions on Graphics (ToG) , volume=. 2021 , publisher=

    2021

  21. [21]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Nerd: Neural reflectance decomposition from image collections , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  22. [22]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , year=

    Nemf: Inverse volume rendering with neural microflake field , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , year=

  23. [23]

    European Conference on Computer Vision , pages=

    NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis , author=. European Conference on Computer Vision , pages=. 2020 , organization=

    2020

  24. [24]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Nerf in the wild: Neural radiance fields for unconstrained photo collections , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  25. [25]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  26. [26]

    ACM Transactions on Graphics (TOG) , volume=

    Practical Inverse Rendering of Textured and Translucent Appearance , author=. ACM Transactions on Graphics (TOG) , volume=. 2025 , publisher=

    2025

  27. [27]

    European Conference on Computer Vision , pages=

    An evaluation of computational imaging techniques for heterogeneous inverse scattering , author=. European Conference on Computer Vision , pages=. 2016 , organization=

    2016

  28. [28]

    ACM Transactions on Graphics (TOG) , volume=

    A differential theory of radiative transfer , author=. ACM Transactions on Graphics (TOG) , volume=. 2019 , publisher=

    2019

  29. [29]

    2020 IEEE International Conference on Computational Photography (ICCP) , pages=

    Towards learning-based inverse subsurface scattering , author=. 2020 IEEE International Conference on Computational Photography (ICCP) , pages=. 2020 , organization=

    2020

  30. [30]

    ACM Transactions on Graphics (TOG) , volume=

    Unbiased inverse volume rendering with differential trackers , author=. ACM Transactions on Graphics (TOG) , volume=. 2022 , publisher=

    2022

  31. [31]

    ACM Transactions on Graphics (TOG) , volume=

    Path-space differentiable rendering of participating media , author=. ACM Transactions on Graphics (TOG) , volume=. 2021 , publisher=

    2021

  32. [32]

    ACM Transactions on Graphics (TOG) , volume=

    Modeling and rendering of heterogeneous translucent materials using the diffusion equation , author=. ACM Transactions on Graphics (TOG) , volume=. 2008 , publisher=

    2008

  33. [33]

    ACM SIGGRAPH 2022 Conference Proceedings , pages=

    Reconstructing translucent objects using differentiable rendering , author=. ACM SIGGRAPH 2022 Conference Proceedings , pages=

    2022

  34. [34]

    SIGGRAPH Asia 2024 Conference Papers , pages=

    Reconstructing translucent thin objects from photos , author=. SIGGRAPH Asia 2024 Conference Papers , pages=

    2024

  35. [35]

    ACM Transactions on Graphics (ToG) , volume=

    Appearance capture and modeling of human teeth , author=. ACM Transactions on Graphics (ToG) , volume=. 2018 , publisher=

    2018

  36. [36]

    Proceedings of the IEEE International Conference on Computer Vision , pages=

    Airborne three-dimensional cloud tomography , author=. Proceedings of the IEEE International Conference on Computer Vision , pages=

  37. [37]

    Computers & Graphics , volume=

    Density estimation for dynamic volumes , author=. Computers & Graphics , volume=. 2007 , publisher=

    2007

  38. [38]

    Computer Graphics Forum , volume=

    BSSRDF estimation from single images , author=. Computer Graphics Forum , volume=. 2011 , organization=

    2011

  39. [39]

    ACM Transactions on Graphics (TOG) , volume=

    Differentiable monte carlo ray tracing through edge sampling , author=. ACM Transactions on Graphics (TOG) , volume=. 2018 , publisher=

    2018

  40. [40]

    ACM Transactions on Graphics (TOG) , volume=

    Reparameterizing discontinuous integrands for differentiable rendering , author=. ACM Transactions on Graphics (TOG) , volume=. 2019 , publisher=

    2019

  41. [41]

    ACM Transactions on Graphics (TOG) , volume=

    Radiative backpropagation: An adjoint method for lightning-fast differentiable rendering , author=. ACM Transactions on Graphics (TOG) , volume=. 2020 , publisher=

    2020

  42. [42]

    ACM Transactions on Graphics (TOG) , volume=

    Monte Carlo estimators for differential light transport , author=. ACM Transactions on Graphics (TOG) , volume=. 2021 , publisher=

    2021

  43. [43]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Compressive structured light for recovering inhomogeneous participating media , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2012 , publisher=

    2012

  44. [44]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  45. [45]

    Advances in Neural Information Processing Systems , volume=

    Neural-pil: Neural pre-integrated lighting for reflectance decomposition , author=. Advances in Neural Information Processing Systems , volume=

  46. [46]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Neural Inverse Rendering from Propagating Light , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  47. [47]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    PBR-NeRF: Inverse Rendering with Physics-Based Neural Fields , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  48. [48]

    European Conference on Computer Vision , pages=

    Neural radiance transfer fields for relightable novel-view synthesis with global illumination , author=. European Conference on Computer Vision , pages=. 2022 , organization=

    2022

  49. [49]

    ACM Transactions on Graphics (TOG) , volume=

    Diffusion posterior illumination for ambiguity-aware inverse rendering , author=. ACM Transactions on Graphics (TOG) , volume=. 2023 , publisher=

    2023

  50. [50]

    European Conference on Computer Vision , pages=

    Nerf for outdoor scene relighting , author=. European Conference on Computer Vision , pages=. 2022 , organization=

    2022

  51. [51]

    ACM SIGGRAPH 2024 Conference Papers , pages=

    Physics-Informed Learning of Characteristic Trajectories for Smoke Reconstruction , author=. ACM SIGGRAPH 2024 Conference Papers , pages=

    2024

  52. [52]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    FluidNexus: 3D fluid reconstruction and prediction from a single video , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  53. [53]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Gs-ir: 3d gaussian splatting for inverse rendering , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  54. [54]

    European Conference on Computer Vision , pages=

    Relightable 3d gaussians: Realistic point cloud relighting with brdf decomposition and ray tracing , author=. European Conference on Computer Vision , pages=

  55. [55]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Relightable gaussian codec avatars , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  56. [56]

    IEEE Transactions on Transactions on Pattern Analysis and Machine Intelligence , year =

    Shi, Yahao and Wu, Yanmin and Wu, Chenming and Liu, Xing and Zhao, Chen and Feng, Haocheng and Zhang, Jian and Zhou, Bin and Ding, Errui and Wang, Jingdong , title =. IEEE Transactions on Transactions on Pattern Analysis and Machine Intelligence , year =

  57. [57]

    2025 , publisher =

    Condor, Jorge and Speierer, Sebastien and Bode, Lukas and Bozic, Aljaz and Green, Simon and Didyk, Piotr and Jarabo, Adrian , title =. 2025 , publisher =. doi:10.1145/3711853 , journal=

    doi:10.1145/3711853 2025

  58. [58]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year =

    Kaleta, Joanna and Kania, Kacper and Trzci. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year =

  59. [59]

    Advances in Neural Information Processing Systems , volume=

    Subsurface Scattering for Gaussian Splatting , author=. Advances in Neural Information Processing Systems , volume=

  60. [60]

    Proceedings of the International Conference on Learning Representations (ICLR) , year =

    GI-GS: Global Illumination Decomposition on Gaussian Splatting for Inverse Rendering , author =. Proceedings of the International Conference on Learning Representations (ICLR) , year =

  61. [61]

    CVPR , year=

    ReCap: Better Gaussian Relighting with Cross-Environment Captures , author=. CVPR , year=

  62. [62]

    arXiv preprint , year=

    Reflective Gaussian Splatting , author=. arXiv preprint , year=

  63. [63]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Tensoir: Tensorial inverse rendering , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  64. [64]

    2023 , publisher=

    Physically based rendering: From theory to implementation , author=. 2023 , publisher=

    2023

  65. [65]

    ACM Transactions on Graphics (TOG) , volume=

    Path replay backpropagation: Differentiating light paths using constant memory and linear time , author=. ACM Transactions on Graphics (TOG) , volume=. 2021 , publisher=

    2021

  66. [66]

    ACM Transactions on Graphics (TOG) , volume=

    Recursive control variates for inverse rendering , author=. ACM Transactions on Graphics (TOG) , volume=. 2023 , publisher=

    2023

  67. [67]

    ACM Transactions on Graphics (ToG) , volume=

    Mitsuba 2: A retargetable forward and inverse renderer , author=. ACM Transactions on Graphics (ToG) , volume=. 2019 , publisher=

    2019

  68. [68]

    Computer graphics forum , volume=

    Neural fields in visual computing and beyond , author=. Computer graphics forum , volume=. 2022 , organization=

    2022

  69. [69]

    ACM Transactions on Graphics (TOG) , volume=

    Neural radiosity , author=. ACM Transactions on Graphics (TOG) , volume=. 2021 , publisher=

    2021

  70. [70]

    ACM SIGGRAPH 2023 Conference Proceedings , pages=

    Inverse global illumination using a neural radiometric prior , author=. ACM SIGGRAPH 2023 Conference Proceedings , pages=

    2023

  71. [71]

    ACM Transactions on Graphics (TOG) , volume=

    Tensosdf: Roughness-aware tensorial representation for robust geometry and material reconstruction , author=. ACM Transactions on Graphics (TOG) , volume=. 2024 , publisher=

    2024

  72. [72]

    European conference on computer vision , pages=

    Neilf: Neural incident light field for physically-based material estimation , author=. European conference on computer vision , pages=. 2022 , organization=

    2022

  73. [73]

    Worchel, Markus and Finnendahl, Ugo and Alexa, Marc , title =. Proc. EGSR , year =

  74. [74]

    Journal of Quantitative Spectroscopy and Radiative Transfer , volume=

    Physics informed neural networks for simulating radiative transfer , author=. Journal of Quantitative Spectroscopy and Radiative Transfer , volume=. 2021 , publisher=

    2021

  75. [75]

    Computer Graphics Forum (Proceedings of the Eurographics Conference) , number =

    Violante, Nicolás and Gauthier, Alban and Diolatzis, Stavros and Leimkühler, Thomas and Drettakis, George , title =. Computer Graphics Forum (Proceedings of the Eurographics Conference) , number =

  76. [76]

    International Conference on 3D Vision (3DV) , year=

    GAN2X: Non-Lambertian Inverse Rendering of Image GANs , author=. International Conference on 3D Vision (3DV) , year=

  77. [77]

    2022 , booktitle=

    Neural Radiance Transfer Fields for Relightable Novel-view Synthesis with Global Illumination , author =. 2022 , booktitle=

    2022

  78. [78]

    Advances in Neural Information Processing Systems , volume=

    A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis , author=. Advances in Neural Information Processing Systems , volume=

  79. [79]

    Computer Graphics Forum , volume=

    Mesogan: Generative neural reflectance shells , author=. Computer Graphics Forum , volume=. 2023 , organization=

    2023

  80. [80]

    ACM Transactions on Graphics , volume=

    Nerffacelighting: Implicit and disentangled face lighting representation leveraging generative prior in neural radiance fields , author=. ACM Transactions on Graphics , volume=. 2023 , publisher=

    2023

Showing first 80 references.