Pith. sign in

REVIEW 4 major objections 4 minor 44 references

Gradient backpropagation through a differentiable renderer yields a saliency map over scene parameters for any scalar metric, exposing which materials, geometries, and lights actually drive the metric.

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

2026-08-01 07:02 UTC pith:FBNAKSAC

load-bearing objection Metric saliency maps are a clean repurposing of differentiable-rendering gradients, honestly presented; the unquantified UGR relaxation is the main soft spot, not a fatal flaw. the 4 major comments →

arxiv 2607.21562 v1 pith:FBNAKSAC submitted 2026-07-23 cs.CV cs.GR

Scene Parameter Saliency via Differentiable Light Transport

Linas Beresna , Eugene Fiume This is my paper
classification cs.CV cs.GR
keywords metric saliencydifferentiable renderinggradient attributionlight transportglobal illuminationUnified Glare Ratingscene understandingsensitivity analysis
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.

The paper establishes that the gradient of any scalar metric with respect to scene parameters, obtained in one reverse-mode pass through a differentiable renderer, is an interpretability tool: it ranks how much each physical parameter (reflectance, roughness, emitter intensity, index of refraction) influences the metric. This matters because it turns a single number, like a discomfort-glare rating, into actionable physical attributions that account for multi-bounce light transport, dependencies a designer cannot read off the rendered image. The paper shows the ranking is metric-specific: the same scene yields different saliency maps for mean luminance, glare, and a neural classifier's logit. It validates the gradient ranking against independent local ablations, confirming that the top-ranked parameters produce the largest measured metric changes. The practical payoff is a cheap local sensitivity analysis at the cost of one backward pass instead of many forward simulations.

Core claim

The central claim is that the parameter-space gradient g = ∂M(R(θ))/∂θ assigns each scene parameter a score equal to how much the metric M would change per unit perturbation of that parameter, through all direct and indirect light paths. A single reverse-mode differentiation through the composed map metric-after-renderer computes this attribution without separately computing image or parameter sensitivities. The paper demonstrates across indoor scenes that these rankings are specific to the metric: a mirror dominates the discomfort-glare saliency but is negligible for mean luminance, which instead concentrates on large diffuse surfaces. In the neural case, the gradient flows through a pretra

What carries the argument

The key mechanism is reverse-mode automatic differentiation applied jointly to a scalar metric and a physically-based renderer with multi-bounce path tracing, factoring as ∂M/∂θ = (∂M/∂I)(∂I/∂θ). This chain-rule product is evaluated in one backward pass, so no repeated simulations are needed. To visualize the parameter-space vector, the paper traces primary rays to build a per-pixel material identity map and colors each pixel by the saliency of the material visible there. For non-differentiable metrics like the Unified Glare Rating, the metric is replaced by a smooth sigmoid relaxation of its discrete glare-source check, which is the step the paper flags as needing further scrutiny.

Load-bearing premise

The load-bearing premise is that the smooth sigmoid relaxation of the discontinuous Unified Glare Rating, introduced to make the metric differentiable, yields gradient rankings that faithfully represent the true UGR's sensitivity; the paper itself notes this relationship is not yet established.

What would settle it

Render a controlled interior scene at fine resolution, compute the metric-saliency ranking from the relaxed UGR, then exhaustively perturb each top parameter by small finite steps and measure the true UGR change. If a parameter with low predicted saliency produces a large true-UGR jump while a high-ranked parameter produces little change, the relaxation has broken the ranking and the psychovisual demonstration would not generalize.

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

If this is right

  • A single backward pass delivers a local sensitivity ranking that previously required dozens of parametric sweeps or variance-based global sensitivity analyses.
  • Saliency maps are metric-specific; the same scene under different objectives points to different physical causes, so the tool answers 'what matters for this criterion' rather than 'what matters in general.'
  • The method applies to any differentiable metric and any differentiable rendering pipeline, including rasterization-based renderers, neural radiance fields, and Gaussian splatting, with multi-bounce settings where it is most valuable.
  • The neural case shows attribution can flow through both learned latent spaces and physical light transport, connecting model interpretability to scene understanding.
  • The validation on glare shows measured local impact follows the predicted ranking monotonically, supporting the use of gradients as a substitute for expensive finite-difference probes.

Where Pith is reading between the lines

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

  • The paper leaves implicit that the same gradient could guide optimization: using the saliency map to decide which parameters to free or fix during inverse rendering, coupling interpretability with automation.
  • An extension the authors do not pursue is using second-order (Hessian) information to capture parameter interactions; a testable intermediate step would be comparing saliency-based parameter selection against variance-based indices on a small scene.
  • The smooth-relaxation caveat suggests a concrete test: on scenes where the true UGR has discontinuities, compare relaxed-gradient rankings against exhaustive quantized sweeps; divergence would show the relaxation limits psychovisual fidelity.
  • Because the saliency is local, it may mislead at boundaries or saturation; averaging the gradient over multiple perturbed states could stabilize rankings, analogous to smoothing methods in neural interpretability.

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

4 major / 4 minor

Summary. The paper proposes 'metric saliency maps': for a differentiable renderer R(θ) and a scalar image metric M, the reverse-mode gradient g = ∂M(R(θ))/∂θ is interpreted as a saliency vector over scene parameters. The authors argue that, like gradient-based saliency in neural networks, this single-pass derivative reveals which scene elements (materials, roughness, index of refraction, lighting) most influence a chosen objective, including through multi-bounce light transport. They demonstrate the idea on three metric families — mean luminance, the CIE Unified Glare Rating (UGR) via a smooth sigmoid relaxation, and a ResNet-50 class logit — and validate the rankings with an independent one-step ablation in two scenes. The central claims are that these maps are metric-specific (the same scene produces qualitatively different rankings for different objectives) and that they expose non-obvious physical drivers, particularly for discomfort glare.

Significance. If the empirical claims are supported, this is a genuinely useful conceptual contribution: it is a cheap, parameter-free way to obtain first-order sensitivity of arbitrary differentiable metrics with respect to scene parameters, and it extends the interpretability toolkit from neural networks to physics-based image formation. The underlying identity in Eq. (1) is definitionally correct, and the method requires no fitted parameters or learned approximations. The ablation-style validation is the right kind of consistency check. However, the current evidence is largely qualitative: metric-specificity is asserted from visual inspection, the UGR relaxation is unspecified, and the validation covers only top-ranked parameters with a single step size and no error bars. The paper's strongest claims — that the maps identify 'true physical drivers of discomfort glare' and that rankings are metric-specific — therefore outrun the quantitative support. With a specified relaxation and a rigorous rank-stability analysis, the paper could be a valuable addition to differentiable rendering and lighting-design literatures.

major comments (4)
  1. [§3.4, §4.3, §5 (Limitations)] The UGR demonstration relies on a smooth sigmoid relaxation of a discontinuous metric, but no formula, temperature, or threshold is reported. Because UGR involves binary glare-source visibility and a discrete sum over sources, the relaxation can create gradient mass in transition regions that is an artifact of the relaxation, not of UGR. The paper itself states in §5 that 'the choice of relaxation can affect the resulting saliency map, and the relationship between the relaxed and the original metric deserves further investigation.' Fig. 5 does not close this gap: if the 'actual response' is evaluated with the same relaxed objective, the monotone agreement only validates the relaxed model. Please provide the relaxation equation, report its parameters, test ranking stability over those parameters, and compare against finite-difference perturbations of the true CIE UGR or a high-accuracy di
  2. [§4.3, Fig. 5] The ablation validation is too narrow to support the ranking claims. It shows only the top-ranked parameters, uses a single step size (0.01), and gives no error bars or repeated-seed statistics. For mean luminance, changes below 10^-4 are admitted to be at the stochastic noise floor, yet no confidence intervals are provided for the UGR case either. Please report the full parameter list, compute rank correlation (e.g., Spearman) between gradient ranks and independently measured |ΔM| at several step sizes and sample counts, and state the noise floor explicitly. Without this, 'perfectly mirrors' and 'minor deviations' are anecdotal rather than quantitative validation.
  3. [§3.2, Eq. (1), §4.3] The saliency magnitude is not invariant under reparameterization: for θ'_i = f(θ_i), one has g'_i = g_i / f'(θ_i). The paper ranks parameters with heterogeneous units and nonlinear parameterizations (reflectance, roughness α, index of refraction η) without defining a canonical scale, and the validation uses a uniform step size across those domains. This makes cross-parameter rankings and the 'true physical drivers' claim sensitive to arbitrary parameter scaling. Please either define a canonical parameterization/normalization, use dimensionless elasticities such as ∂log M / ∂log θ, or restrict ranking claims to within-parameter-type comparisons.
  4. [§4.1, Fig. 3] The paper's third contribution is that 'saliency rankings differ substantially across metrics'; the evidence, however, is a qualitative side-by-side of heatmaps and bar charts. Please quantify metric specificity: report rank correlation or distributional divergence between saliency vectors for mean luminance, UGR, and the neural logit on the same scene, and across scenes. This would also establish that metric specificity is not an artifact of the visualization normalization described in §3.3.
minor comments (4)
  1. [§4, §5] The experimental setup states 1024 samples per pixel, while the Limitations section says 'moderate sample counts (64–256 spp) produced stable rankings.' Clarify which sample count was used for each reported result.
  2. [§3.4, §4.2] The 'neural perceptual score' is described in §3.4 and §4.2 as a pre-softmax class logit of a ResNet-50 classifier. This is not a perceptual similarity score; rename it to 'class logit' or 'neural classification score' to avoid confusion.
  3. [§3.3] The image-space visualization colors each pixel by the saliency of the material directly visible at that pixel, but a parameter's gradient is global: a high-saliency parameter on a surface not directly visible will be absent from the map. State this approximation explicitly and note that the map is an object-level visualization, not a per-pixel derivative field.
  4. [Fig. 5] Please define the exact physical meaning and units of the parameters labeled α, k, and η, and state the parameter ranges used; this is needed to interpret the uniform step size and the cross-parameter rankings.

Circularity Check

1 steps flagged

UGR validation reduces to checking the gradient against itself; relaxation gap remains untested.

specific steps
  1. self definitional [§4.3 Validation via Independent Parameter Ablation, with Eq. (1) from §3.2 and §5 Limitations]
    "To verify that our computed gradients accurately predict the physical sensitivities of the objective functions, we perform an independent ablation test on both the non-linear UGR metric and the global mean luminance metric (Figure 5). ... we evaluate the metric’s actual response to a small, independent gradient descent step (learning rate=0.01) for each of the top-ranked parameters. ... As shown in Figure 5a, the absolute measured impact on the localised UGR metric perfectly follows our computed saliency ranking in a monotonically decreasing curve. This confirms that a single reverse-mode eval"

    The 'actual response' is evaluated for the same differentiable objective whose gradient defines the saliency ranking. For a parameter step δθ_i = η g_i (or η sign(g_i)), the first-order change is |ΔM| ≈ |g_i δθ_i| = η g_i² (or η|g_i|), so the ranking by |g_i| and the ranking by measured |ΔM| coincide by Eq. (1) for any smooth M. The monotonic curve in Fig. 5a is therefore a consistency check of the chain rule, not an independent confirmation that the relaxed-UGR gradients correspond to the true UGR. The paper's own limitation, 'The choice of relaxation can affect the resulting saliency map, and the relationship between the relaxed and the original metric deserves further investigation,' concedes this gap, so the claim that the method 'ranks the true physical drivers of discomfort glare' is

full rationale

The core method—computing ∂M(R(θ))/∂θ by reverse-mode differentiation—is not circular: Eq. (1) is the chain rule, and for a differentiable metric the gradient is by definition the local sensitivity. The metric-specificity observation is also a direct consequence of gradients being function-specific, but that is a tautology rather than a circular derivation. The significant circularity is in the validation of the headline UGR demonstration. The ablation checks whether the gradient ranking predicts the actual response of the objective to a small step; for the same differentiable objective, this is forced to first order by Taylor's theorem, so the 'perfectly' monotonic curve in Fig. 5a is guaranteed by construction. Since the paper uses a smooth sigmoid relaxation of UGR and explicitly states that the relationship between the relaxed and original metric is uninvestigated, the validation cannot support the claim that the saliency map ranks the true physical drivers of discomfort glare. This is a partial circularity: the prediction reduces to the definition of the derivative of the surrogate objective, not to an independent test of UGR sensitivity. No self-citation load-bearing issues or fitted-input-as-prediction problems appear elsewhere.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The method itself is cheap: it adds no physical entities and fits no global constants. The load-bearing choices are the smoothness of the metric (UGR relaxation), the comparability of gradients across parameter types, and the trustworthiness of the path-tracing gradient estimator. These are the places where the paper's conclusions are most fragile.

free parameters (3)
  • UGR smooth sigmoid relaxation parameters = not reported
    UGR's binary glare-source visibility must be smoothed to be differentiable (§5 Limitations). The paper does not report the relaxation's scale or sharpness and states it can change the saliency map, so the maps depend on a hand-chosen, unreported smoothing.
  • Validation perturbation step size = 0.01
    Used in the independent ablation (Fig. 5). The cross-domain step (reflectance vs eta) is acknowledged to cause nonuniform behavior, so the validation ranking depends on this arbitrary step.
  • Rendering sample count and max path depth = 1024 spp, depth 6
    Chosen for the experiments (§4). Gradient estimates inherit MC variance; at lower spp rankings become unstable as the paper notes, so the demonstrated stable rankings depend on these hand-set budgets.
axioms (4)
  • domain assumption Reverse-mode gradients through Mitsuba 3 with Path Replay Backpropagation provide unbiased estimates of ∂I/∂θ including multi-bounce transport.
    Invoked in §3.1; the entire method rests on the correctness of the differentiable path tracer's gradient estimator.
  • ad hoc to paper The sigmoid-smoothed version of UGR has a gradient whose ranking faithfully represents the true UGR's sensitivity.
    Introduced in §5 Limitations: 'The choice of relaxation can affect the resulting saliency map'; used to make a discontinuous metric differentiable.
  • domain assumption Gradient magnitudes |g_i| are directly comparable across parameters with different physical units and scales.
    The bar charts in Fig. 5 rank reflectance, roughness (alpha), and index of refraction together; the paper only notes in future work that cross-domain step scaling is nonlinear, so ranking validity is assumed.
  • domain assumption A per-pixel primary-ray material identity map is sufficient to visualize parameter saliency in image space.
    Used in §3.3; holds for uniform material parameters, explicitly deferred for spatially-varying textures.

pith-pipeline@v1.3.0-alltime-deepseek · 9516 in / 14161 out tokens · 139493 ms · 2026-08-01T07:02:58.703304+00:00 · methodology

0 comments
read the original abstract

Gradient-based saliency methods reveal which input features most influence a neural network's output, and are a standard tool for model interpretability. We observe that differentiable renderers, which are conventionally used for parameter optimisation, produce an analogous form of saliency: given any scalar metric evaluated on a rendered image, a single reverse-mode differentiation pass yields per-parameter gradients that identify which scene elements most influence the metric. We call these gradient fields metric saliency maps. Unlike neural saliency, which propagates attribution through learned weights, metric saliency propagates through the image formation process itself, including multi-bounce light transport, capturing parameter dependencies that are semi-opaque to manual inspection. We compute metric saliency maps for qualitatively different objectives: psychovisual glare indices, mean scene luminance, and neural perceptual scores. The saliency rankings differ substantially across metrics for the same scene, with parameters that dominate one objective being negligible for another. The saliency map is specific to the metric, not an intrinsic property of the scene. Our results suggest that differentiable renderers produce derivative images that are as informative for scene understanding as the primal images they were designed to generate.

Figures

Figures reproduced from arXiv: 2607.21562 by Eugene Fiume, Linas Beresna.

Figure 1
Figure 1. Figure 1: Expanding the diagnostic toolkit for architectural lighting. While conven￾tional analysis provides (b) glare source identification and (c) luminance maps to evaluate discomfort glare, it leaves the designer to guess the root physical causes. By differentiating the psychovisual unified glare rating through the light transport simulation, we obtain (d) a metric saliency map. This highlights the specific scen… view at source ↗
Figure 2
Figure 2. Figure 2: Image-Space Saliency Map￾ping. Per-pixel material identities (Scene Objects) are coloured by their corre￾sponding parameter sensitivities (Scene Gradients) to produce the final metric￾specific heatmap. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualizing Metric Specificity. While extracting gradients provides precise rank￾ing data for scene parameters (top row), mapping this raw data back into the 3D scene (bot￾tom row) allows for intuitive spatial visualisation. Across multiple scenes, the UGR evalua￾tion isolates highly localised reflective surfaces driving discomfort glare (e.g., the mirror or metallic utensils). In contrast, the mean lumina… view at source ↗
Figure 4
Figure 4. Figure 4: Semantic attribution mapped into 3D parameter space. The saliency map shows which materials act as the cues for the classification logit. We demonstrate that this pipeline handles high dimensional learned spaces by evaluating a se￾mantic classifier [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Independent Ablation Validation. Parameters are ranked strictly by computed gradient magnitude (y-axis). Taking a small gradient descent step yields an absolute change in the objective metric (x-axis). The UGR evaluation (a) perfectly mirrors the computed saliency rank, whilst the mean luminance evaluation (b) highlights the noise floor challenges of global stochastic metrics. As shown in Figure 5a, the ab… 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

44 extracted references · 4 linked inside Pith

  1. [1]

    Shape, illumination, and reflectance from shad- ing.IEEE transactions on pattern analysis and machine intelligence, 37(8):1670–1687, 2014

    Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shad- ing.IEEE transactions on pattern analysis and machine intelligence, 37(8):1670–1687, 2014

    2014

  2. [2]

    Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin- Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. InProceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021

    2021

  3. [3]

    Recovering intrinsic scene characteristics.Comput

    Harry Barrow, J Tenenbaum, A Hanson, and E Riseman. Recovering intrinsic scene characteristics.Comput. vis. syst, 2(3-26):2, 1978

    1978

  4. [4]

    Mrd: Using physically based differen- tiable rendering to probe vision models for 3d scene understanding.arXiv preprint arXiv:2512.12307, 2025

    Benjamin Beilharz and Thomas SA Wallis. Mrd: Using physically based differen- tiable rendering to probe vision models for 3d scene understanding.arXiv preprint arXiv:2512.12307, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Rendering resources, 2016

    Benedikt Bitterli. Rendering resources, 2016. https://benedikt-bitterli.me/resources/

    2016

  6. [6]

    Saliency cards: A framework to characterize and compare saliency meth- ods

    Angie Boggust, Harini Suresh, Hendrik Strobelt, John Guttag, and Arvind Satya- narayan. Saliency cards: A framework to characterize and compare saliency meth- ods. InProceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 285–296, 2023. LINAS BERESNA AND EUGENE FIUME: SCENE PARAMETER SALIENCY11

    2023

  7. [7]

    Grad-cam++: Generalized gradient-based visual explanations for deep convolu- tional networks

    Aditya Chattopadhay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubrama- nian. Grad-cam++: Generalized gradient-based visual explanations for deep convolu- tional networks. In2018 IEEE winter conference on applications of computer vision (WACV), pages 839–847. IEEE, 2018

    2018

  8. [8]

    Learning to predict 3d objects with an interpolation-based differen- tiable renderer.Advances in neural information processing systems, 32, 2019

    Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differen- tiable renderer.Advances in neural information processing systems, 32, 2019

    2019

  9. [9]

    Fully automatic id mattes with support for motion blur and transparency

    Jonah Friedman and Andrew C Jones. Fully automatic id mattes with support for motion blur and transparency. InACM SIGGRAPH 2015 Posters, pages 1–1. 2015

    2015

  10. [10]

    Relightable 3d gaussians: Realistic point cloud relighting with brdf decomposition and ray tracing

    Jian Gao, Chun Gu, Youtian Lin, Zhihao Li, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussians: Realistic point cloud relighting with brdf decomposition and ray tracing. InEuropean Conference on Computer Vision, pages 73–89. Springer, 2024

    2024

  11. [11]

    Computing visual comfort ratings for a specific interior lighting installation.Illuminating Engineering, 61(10):634, 1966

    Sylvester K Guth. Computing visual comfort ratings for a specific interior lighting installation.Illuminating Engineering, 61(10):634, 1966

    1966

  12. [12]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  13. [13]

    Daylighting and human performance.ASHRAE journal, 44(6):65–67, 2002

    Lisa Heschong. Daylighting and human performance.ASHRAE journal, 44(6):65–67, 2002

    2002

  14. [14]

    Berthold K. P. Horn and Michael J. Brooks, editors.Shape from shading. MIT Press, Cambridge, MA, USA, 1989. ISBN 0262081830

    1989

  15. [15]

    Discomfort glare in interior lighting

    International Commission on Illumination. Discomfort glare in interior lighting. Tech- nical Report CIE 117-1995, CIE, Vienna, Austria, 1995. ISBN 978-3-900734-70-1

    1995

  16. [16]

    Mitsuba 3 renderer, 2022

    Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org

    2022

  17. [17]

    Dr.jit: A just-in- time compiler for differentiable rendering.Transactions on Graphics (Proceedings of SIGGRAPH), 41(4), July 2022

    Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, and Delio Vicini. Dr.jit: A just-in- time compiler for differentiable rendering.Transactions on Graphics (Proceedings of SIGGRAPH), 41(4), July 2022. doi: 10.1145/3528223.3530099

    arXiv 2022

  18. [18]

    Neural 3d mesh renderer

    Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3907–3916, 2018

    2018

  19. [19]

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

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, George Drettakis, et al. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4): 139–1, 2023

    2023

  20. [20]

    Modular primitives for high-performance differentiable rendering.ACM Trans- actions on Graphics (ToG), 39(6):1–14, 2020

    Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering.ACM Trans- actions on Graphics (ToG), 39(6):1–14, 2020. 12LINAS BERESNA AND EUGENE FIUME: SCENE PARAMETER SALIENCY

    2020

  21. [21]

    Differentiable monte carlo ray tracing through edge sampling.ACM Transactions on Graphics (TOG), 37 (6):1–11, 2018

    Tzu-Mao Li, Miika Aittala, Frédo Durand, and Jaakko Lehtinen. Differentiable monte carlo ray tracing through edge sampling.ACM Transactions on Graphics (TOG), 37 (6):1–11, 2018

    2018

  22. [22]

    View-independent adjoint light tracing for lighting design optimization.ACM Trans

    Lukas Lipp, David Hahn, Pierre Ecormier-Nocca, Florian Rist, and Michael Wimmer. View-independent adjoint light tracing for lighting design optimization.ACM Trans. Graph., 43(3), May 2024. ISSN 0730-0301. doi: 10.1145/3662180

    doi:10.1145/3662180 2024

  23. [23]

    Soft rasterizer: A differentiable renderer for image-based 3d reasoning

    Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. InProceedings of the IEEE/CVF international conference on computer vision, pages 7708–7717, 2019

    2019

  24. [24]

    Approved method: Ies spatial daylight autonomy (sda) and annual sunlight exposure (ase).Illuminating Engineering Society

    I Lm. Approved method: Ies spatial daylight autonomy (sda) and annual sunlight exposure (ase).Illuminating Engineering Society. https://www. ies. org/product/ies- spatial-daylight-autonomy-sda-and-annual-sunlight-exposure-ase, 2013

    2013

  25. [25]

    Reparameterizing discon- tinuous integrands for differentiable rendering.Transactions on Graphics (Proceedings of SIGGRAPH Asia), 38(6), December 2019

    Guillaume Loubet, Nicolas Holzschuch, and Wenzel Jakob. Reparameterizing discon- tinuous integrands for differentiable rendering.Transactions on Graphics (Proceedings of SIGGRAPH Asia), 38(6), December 2019. doi: 10.1145/3355089.3356510

    arXiv 2019

  26. [26]

    Nerf: Representing scenes as neural radiance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ra- mamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021

    2021

  27. [27]

    Factorial sampling plans for preliminary computational experiments

    Max D Morris. Factorial sampling plans for preliminary computational experiments. Technometrics, 33(2):161–174, 1991

    1991

  28. [28]

    Useful daylight illuminances: A replacement for daylight factors.Energy and buildings, 38(7):905–913, 2006

    Azza Nabil and John Mardaljevic. Useful daylight illuminances: A replacement for daylight factors.Energy and buildings, 38(7):905–913, 2006

    2006

  29. [29]

    Mitsuba 2: A retargetable forward and inverse renderer.Transactions on Graphics (Proceedings of SIGGRAPH Asia), 38(6), December 2019

    Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer.Transactions on Graphics (Proceedings of SIGGRAPH Asia), 38(6), December 2019. doi: 10.1145/3355089.3356498

    arXiv 2019

  30. [30]

    Ra- diative backpropagation: An adjoint method for lightning-fast differentiable render- ing.Transactions on Graphics (Proceedings of SIGGRAPH), 39(4), July 2020

    Merlin Nimier-David, Sébastien Speierer, Benoît Ruiz, and Wenzel Jakob. Ra- diative backpropagation: An adjoint method for lightning-fast differentiable render- ing.Transactions on Graphics (Proceedings of SIGGRAPH), 39(4), July 2020. doi: 10.1145/3386569.3392406

    arXiv 2020

  31. [31]

    Daylight performance predictions

    Christoph Reinhart. Daylight performance predictions. InBuilding performance simu- lation for design and operation, pages 221–269. Routledge, 2019

    2019

  32. [32]

    Differentiable objectives for 3d scene relighting via gradient descent on olat basis coefficients

    Anson Savage. Differentiable objectives for 3d scene relighting via gradient descent on olat basis coefficients. 2026

    2026

  33. [33]

    Grad-cam: Visual explanations from deep networks via gradient-based localization

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. InProceedings of the IEEE international conference on computer vision, pages 618–626, 2017

    2017

  34. [34]

    Deep inside convolutional networks: Visualising image classification models and saliency maps.arXiv preprint arXiv:1312.6034, 2013

    Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps.arXiv preprint arXiv:1312.6034, 2013. LINAS BERESNA AND EUGENE FIUME: SCENE PARAMETER SALIENCY13

    Pith/arXiv arXiv 2013

  35. [35]

    Smoothgrad: removing noise by adding noise.arXiv preprint arXiv:1706.03825, 2017

    Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise.arXiv preprint arXiv:1706.03825, 2017

    Pith/arXiv arXiv 2017

  36. [36]

    Global sensitivity indices for nonlinear mathematical models and their monte carlo estimates.Mathematics and computers in simulation, 55(1-3):271–280, 2001

    Ilya M Sobol. Global sensitivity indices for nonlinear mathematical models and their monte carlo estimates.Mathematics and computers in simulation, 55(1-3):271–280, 2001

    2001

  37. [37]

    Axiomatic attribution for deep net- works

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep net- works. InInternational conference on machine learning, pages 3319–3328. PMLR, 2017

    2017

  38. [38]

    Diff-dope: Differentiable deep object pose estimation

    Jonathan Tremblay, Bowen Wen, Valts Blukis, Balakumar Sundaralingam, Stephen Tyree, and Stan Birchfield. Diff-dope: Differentiable deep object pose estimation. arXiv preprint arXiv:2310.00463, 2023

    Pith/arXiv arXiv 2023

  39. [39]

    Path replay backpropagation: Differentiating light paths using constant memory and linear time.Transactions on Graphics (Proceedings of SIGGRAPH), 40(4):108:1–108:14, August 2021

    Delio Vicini, Sébastien Speierer, and Wenzel Jakob. Path replay backpropagation: Differentiating light paths using constant memory and linear time.Transactions on Graphics (Proceedings of SIGGRAPH), 40(4):108:1–108:14, August 2021. doi: 10.1145/3450626.3459804

    arXiv 2021

  40. [40]

    Stochastic gradient estimation for higher-order differentiable rendering

    Zican Wang, Michael Fischer, and Tobias Ritschel. Stochastic gradient estimation for higher-order differentiable rendering. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 28198–28206, 2025

    2025

  41. [41]

    Evaluation methods and development of a new glare prediction model for daylight environments with the use of ccd cameras.Energy and buildings, 38(7):743–757, 2006

    Jan Wienold and Jens Christoffersen. Evaluation methods and development of a new glare prediction model for daylight environments with the use of ccd cameras.Energy and buildings, 38(7):743–757, 2006

    2006

  42. [42]

    Photometric method for determining surface orientation from mul- tiple images.Optical engineering, 19(1):139–144, 1980

    Robert J Woodham. Photometric method for determining surface orientation from mul- tiple images.Optical engineering, 19(1):139–144, 1980

    1980

  43. [43]

    Diffcsg: Differentiable csg via rasterization

    Haocheng Yuan, Adrien Bousseau, Hao Pan, Quancheng Zhang, Niloy J Mitra, and Changjian Li. Diffcsg: Differentiable csg via rasterization. InSIGGRAPH Asia 2024 Conference Papers, pages 1–10, 2024

    2024

  44. [44]

    Vision as bayesian inference: Analysis by synthesis? Trends in cognitive sciences, 10(7):301–308, 2006

    Alan Yuille and Daniel Kersten. Vision as bayesian inference: Analysis by synthesis? Trends in cognitive sciences, 10(7):301–308, 2006

    2006