Pith. sign in

REVIEW 2 major objections 6 minor 53 references

Lighting in Motion: Spatiotemporal HDR Lighting Estimation

T0 review · 2 major / 6 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 LiMo predicts full HDR lighting at any 3D point in a scene, at any time in a video, using diffusion models conditioned on geometric maps that ground the prediction in space.

desk verdict Real advance in HDR lighting estimation, but the spatial-grounding claim needs a depth/FOV sensitivity test before I'd sign off. read the letter →

arxiv 2512.13597 v2 pith:SR5YOBH2 submitted 2025-12-15 cs.CV

Christophe Bolduc , Julien Philip , Li Ma , Mingming He , Paul Debevec , Jean-François Lalonde This is my paper
classification cs.CV
keywords spatiotemporallightingestimationHDRdiffusionmodelHDRIspatially-varyinglightprobesinverserenderingvirtualproduction
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

This paper claims the first single framework that estimates full HDR lighting from a monocular image or video at any user-specified 3D point and time, without physical light probes. It argues that prior methods trade off spatial grounding against physical accuracy, and that a diffusion model conditioned on new geometric relation maps — not just depth — can do both. The approach predicts mirror and diffuse spheres at multiple exposures, then fuses them into a single HDRI map through differentiable rendering. If correct, this would let artists insert virtual objects into video with shading that matches the local environment, indoors and outdoors, without visiting the set.

What carries the argument

The two load-bearing components are the geometric conditioning and the HDRI fusion. For each pixel, a direction map and a distance map express, in camera coordinates, where the scene point behind that pixel sits relative to the sphere center; these maps, along with RGB, depth, sphere normals, and an exposure/material text prompt, are concatenated as conditioning for a fine-tuned diffusion model that inpaints the sphere. On the reconstruction side, an equirectangular HDRI represented as a Laplacian pyramid in log space is optimized with a differentiable Monte Carlo renderer — with mirror and diffuse modes, a saturation mask, and a temporal-smoothness loss — against the stack of predicted sphe

What would settle it

Render a controlled scene with known lighting and a probe at a known 3D position, then run LiMo twice: once with the true depth map and once with depth perturbed by realistic noise (for example, a 5 percent bias or blur). If the predicted HDRI changes materially with the depth perturbation, or if using ground-truth depth at test time yields clearly better spatial-grounding metrics than the paper reports, the claim of robust spatial grounding is weakened.

Watch

Extended reading notes

Core claim

LiMo is a diffusion-based lighting estimator that, given an image or video and a target 3D position, inpaints a mirror or diffuse sphere at that position. Its central claim is that two geometric conditioning maps — the direction from each scene point to the target point, and the distance to it — anchor the prediction in 3D space in a way depth alone cannot, so the model correctly distinguishes a sphere in shadow from one in sunlight even when they occupy identical screen-space areas. Querying the model at several exposure levels and two materials, then optimizing a single HDRI against all predictions with a differentiable renderer, yields maps that are simultaneously accurate in luminance an

Load-bearing premise

The spatial-grounding claim depends on off-the-shelf depth and field-of-view estimators being accurate enough at test time; the model is trained on ground-truth values, and the paper does not test what happens when those estimates are noisy.

Editorial extensions

If this is right

  • Virtual objects inserted into video receive locally correct lighting: a probe placed in shadow predicts a dark environment, a probe in sunlight predicts direct light, even at identical screen size.
  • Lighting can be queried at any 3D point in a frame, so a compositor can place an object anywhere and get a matching HDRI rather than a single global estimate.
  • The video version of the model produces temporally stable predictions that track moving cameras, moving objects, and changing light sources, reducing flicker in relit sequences.
  • The output is an explicit HDRI map, so it plugs into the image-based lighting pipelines already used in film and games, preserving artistic control.
  • One trained model covers indoor and outdoor scenes and both near-field and distant light, so a single system can serve many production shot types.

Reading between the lines

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

  • The method's spatial grounding is only as good as the off-the-shelf depth and field-of-view estimators used at test time; the paper trains on ground-truth values and does not ablate sensitivity to estimator noise, so real-world errors in either could shift the anchor of the predicted lighting.
  • The geometric-relation map idea is transferable: any spatially-varying prediction task where depth cues are ambiguous — relighting, inverse rendering, shadow prediction — could benefit from explicitly encoding direction and distance to a target point.
  • The multi-exposure diffuse-plus-mirror fusion is a recipe that could be added to other diffusion-based lighting estimators to improve color and luminance accuracy even where spatial grounding is less critical.
  • A point-based lighting representation, rather than a sphere of finite radius, would remove the paper's stated ill-posedness when a shadow falls on the sphere, suggesting a natural next step.
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

2 major / 6 minor

Summary. The paper proposes LIMO, a diffusion-based method for spatiotemporal HDR lighting estimation. Given a single image or video and a specified 3D scene position, LIMO fine-tunes pretrained image and video diffusion models (Flux.1 Schnell and Wan2.2) to inpaint mirror and diffuse spheres at multiple exposures. The key conditioning novelty is a pair of geometric maps, I_dir and I_dist (Eqs. 1–2, Sec. 3.2), which encode the direction from each scene point to the sphere and the distance to the sphere, computed from per-pixel depth and camera FOV. At inference, the predicted sphere stacks are merged into an equirectangular HDRI by optimizing a differentiable renderer against the network predictions. The method is evaluated on Infinigen Indoor and Laval Indoor SV single-image datasets against DiffusionLight and 4D Lighting, on a synthetic video benchmark built from Blender demo files, and with ablations of the geometric maps and the diffuse sphere. The paper claims to be the first single framework to combine spatial grounding, temporal adaptation, accurate HDR luminance, near/far-field lighting, and high-frequency detail, and claims state-of-the-art results on the evaluated benchmarks.

Significance. If the results hold, LIMO is a meaningful advance for spatially-varying HDR lighting estimation. The reported improvements are large, especially on diffuse, glossy, and matte relit-sphere metrics, and the design is well motivated: diffuse spheres integrate high-energy sources and make luminance estimation more stable, while the geometric maps provide a physically grounded way to condition on a 3D location. The paper also contributes a large synthetic training dataset, clear ablations showing that both the geometric maps and the diffuse sphere term are important, and evaluation against public implementations of two strong baselines. The main significance is conditional on two evaluation gaps: the method's sensitivity to errors in off-the-shelf depth and FOV estimators is not measured, and the video/temporal claims are validated only on synthetic data rendered in the same ecosystem as the training data.

major comments (2)
  1. [Sec. 3.2, Eqs. (1)–(2); Sec. 4.5] The geometric maps I_dir and I_dist are computed from per-pixel depth and camera FOV, using ground-truth depth/FOV at training time but pretrained estimators (Chen et al. [5], Wang et al. [43]) at test time. No experiment measures how errors in these estimators propagate through the unprojection of p_i and the sphere center c. Since Table 3 shows that removing the geometric maps substantially degrades performance (e.g., diffuse RMSE 0.229 vs 0.160), these maps are load-bearing for the central spatial-grounding claim. A biased depth or FOV estimate will corrupt I_dir/I_dist and can also misplace the projected sphere footprint, potentially invalidating the 'grounding at a specified 3D point' claim on real inputs. Please add an ablation on Infinigen (where ground-truth depth/FOV are available) comparing GT vs estimated depth/FOV, report the resulting metrics and sphere-placement error, and
  2. [Sec. 4.2 and Sec. 4.4, Tables 2 and 4] The video evaluation is entirely synthetic: five Blender demo files are augmented with moving cameras, moving probes, and dynamic lighting. The training data are also Blender/BlenderKit renders, so the video test set shares the rendering engine and a similar asset/distribution with training. The only real-data evaluation (Laval Indoor SV, Table 1) is single-image only. Consequently, the temporal and video components of the five-capability claim are not externally validated. The statement that LIMO establishes state-of-the-art for 'spatial control and prediction accuracy' should either be restricted to the single-image setting or supported by a real-video benchmark with captured HDR probes; at minimum, real-video qualitative results should be presented with an explicit caveat that no real-video ground-truth evaluation is available.
minor comments (6)
  1. [Sec. 4.4, Tables 2/4] Please report error bars or significance tests. With 28 Infinigen scenes and 5 video sequences, several reported differences are small (e.g., T-LPIPS-Diff for dynamic lighting diffuse is 0.0005 for both 4D Lighting and LIMO-video). The statement that LIMO 'outperforms' in all cases is stronger than the evidence supports without variance estimates.
  2. [Sec. 4.4] The text notes that 4D Lighting often has lower Warped Err and attributes this to over-smoothing from its MLP formulation. This is plausible but not directly supported by the reported metrics; a perceptual comparison or a metric that penalizes temporal under-variation would strengthen the argument.
  3. [Eqs. (2), (5), and Sec. 3.4] Notation should be cleaned: Eq. (2) uses c_i while Eq. (1) uses c; Eq. (5) should make explicit that the optimization is over the sequence L_t (not a single L), and the loss argument should be written as a comparison, e.g., ell(pi(e,m,t), e R(L_t,m)). Also, 'E=2{EV0,EV−3,...}' in Sec. 3.4 appears to contain a stray '2'.
  4. [Sec. 3.2 and Fig. 2] The test-time input interface for the 3D position should be clarified: is the user specifying a 3D coordinate, clicking a pixel and unprojecting with estimated depth, or using another mechanism? This is important for reproducibility and for assessing the impact of depth/FOV estimation errors.
  5. [Sec. 2 and Sec. 4.2] The claim of state-of-the-art should acknowledge and, if possible, compare against the concurrent LuxDiT [28], which is cited but not included in the baselines. If the public implementation is unavailable, this should be stated.
  6. [Sec. 5] The limitations list is honest about sphere-shadow and near-object cases, but it omits the depth/FOV-estimator sensitivity and the synthetic-only video validation discussed above. Adding these caveats would improve the paper's accuracy.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; central derivation is self-contained, with only a minor non-load-bearing self-citation.

full rationale

LiMo's derivation chain is self-contained. The geometric condition maps (Eqs. 1–2) are computed from depth, camera FOV, and the target 3D position; the diffusion model is trained on rendered synthetic spheres; and the final HDRI is recovered by optimizing Eq. 5 so that a differentiable renderer reproduces the network's multi-exposure mirror/diffuse sphere predictions. The only step that might look self-referential is this HDRI optimization, because the optimized map is fitted to the network's own sphere outputs. That is a test-time inverse-rendering reconstruction loop, not a circular derivation: the objective is not the evaluation metric, and the reported accuracy is externally grounded by relighting held-out spheres from Infinigen Indoor, the Laval Indoor SV dataset, and the Blender-based video test set, including semi-rough/glossy materials that are not part of the Eq. 5 optimization. The exposure-conditioning mention 'Similar to Bolduc et al. [3]' is a minor methodological self-citation and is not load-bearing; no uniqueness theorem is imported, no fitted parameter is renamed as a prediction, and no equation reduces by construction to its inputs. Thus no circularity step rises to the level of the enumerated patterns; the paper's comparatively weaker point is the unablated sensitivity to test-time depth/FOV estimators, which is a robustness concern, not circular reasoning.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper contributes a method rather than a theory; it has no new derived constants. Its empirical claims rest on a small number of hand-chosen hyperparameters (EV brackets, lambda, tau, depth sampling), none of which are swept, and on synthetic-to-real transfer assumptions that are only partially tested.

free parameters (5)
  • Exposure brackets E = {0,-3,-6,-9,-12} EV
    Hand-chosen set of exposures that defines the dynamic range and number of network queries; the HDRI reconstruction can only represent radiance within the bracket coverage.
  • Temporal loss weight lambda = 0.1
    Hand-chosen in Eq. (6) to balance per-frame fidelity against temporal smoothing; no sensitivity study is reported.
  • Sphere depth sampling parameters dmin, dmax, alpha = 0.25, 0.98, 0.4
    Hand-chosen (Eq. 3) to set the training distribution of sphere depths; affects how often the model sees close vs far probes.
  • Saturation mask threshold tau = unspecified
    Appears in Eq. (9) of the supplementary but no numeric value is given; it controls which pixels are excluded from the HDRI optimization loss.
  • HDRI optimization schedule = 1000 iterations/frame, Adam LR 5e-3, 8-level Laplacian pyramid at 512x256
    Hand-chosen convergence settings; no ablation ensures the optimum is reached.
assumptions (4)
  • domain assumption Test-time depth and FOV estimators are accurate enough for the geometric maps.
    Sec. 3.2 uses Chen et al. [5] and Wang et al. [43] at test time while training with ground truth; no ablation.
  • standard math The differentiable renderer faithfully models mirror and diffuse sphere appearance.
    Sec. 7 (supplement): reflection equation (10) and importance-sampled cosine rendering (15); the same physical model is used to generate training targets, so this is internal consistency, not independent validation.
  • domain assumption Synthetic Blender/BlenderKit data transfers to real scenes and to Blender-demo video tests.
    Secs. 3.3/4.4: training and video test sets are Blender-family renders; only single-image generalization is checked on the real Laval dataset.
  • domain assumption Diffusion priors fine-tuned on the custom synthetic domain retain enough photorealism for HDR regressions.
    Sec. 4.1: fine-tuning Flux.1 Schnell and Wan2.2 on 12,896 images / 30,096 sequences assumes the priors' world knowledge survives fine-tuning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Lighting in Motion: Spatiotemporal HDR Lighting Estimation." pith.science (2026). https://pith.science/paper/SR5YOBH2

@misc{pith2026251213597,
  author       = {Pith},
  title        = {Pith review of: Lighting in Motion: Spatiotemporal HDR Lighting Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SR5YOBH2}},
  note         = {Machine review of arXiv:2512.13597}
}
read the original abstract

We present Lighting in Motion (LiMo), a diffusion-based approach to spatiotemporal lighting estimation. LiMo targets both realistic high-frequency detail prediction and accurate illuminance estimation. To account for both, we propose generating a set of mirrored and diffuse spheres at different exposures, based on their 3D positions in the input. Making use of diffusion priors, we fine-tune powerful existing diffusion models on a large-scale customized dataset of indoor and outdoor scenes, paired with spatiotemporal light probes. For accurate spatial conditioning, we demonstrate that depth alone is insufficient and we introduce a new geometric condition to provide the relative position of the scene to the target 3D position. Finally, we combine diffuse and mirror predictions at different exposures into a single HDRI map leveraging differentiable rendering. We thoroughly evaluate our method and design choices to establish LiMo as state-of-the-art for both spatial control and prediction accuracy.

Figures

Figures reproduced from arXiv: 2512.13597 by the authors.

Figure 1
Figure 1. We present LIMO: a spatiotemporal lighting estimation method with accurate spatial grounding, full HDR and realistic reflections. LIMO accurately grounds virtual objects over different spatial positions (1st and 2nd left), and over time (3rd left). LIMO can readily be used in virtual production (right), for instance by inserting actors captured in light domes in real sets. Abstract We present Lighting in Motion (LIM… view at source ↗
Figure 2
Figure 2. Overview of LIMO, our proposed diffusion-based spatiotemporal lighting estimation method. From an input image (or video sequence) and a 3D scene point, we first obtain an estimate of the per-pixel depth (1, top). From this, a set of condition maps are computed (1, bottom). These maps, along with a text prompt, are used to condition a diffusion model (2) which is trained to inpaint a sphere at the desired 3D scene po… view at source ↗
Figure 3
Figure 3. Sample frames from the training dataset, illustrating the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Sample predictions from the Infinigen test set [ [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Example qualitative prediction results on our proposed video test set. Observe how our predictions are more detailed and more [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The effect of our proposed geometric maps [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Additional sample predictions from the Laval Indoor Spatially Varying test set [ [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Additional examples of our method on in-the-wild images and videos, with from left to right: the input frame, the predicted mirror [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 2 linked inside Pith

  1. [5]

    Video depth anything: Consistent depth estimation for super-long videos

    Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zi- long Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos. InIEEE/CVF Conf. Comput. Vis. Pattern Recog., 2025. 2, 4

    2025

  2. [43]

    MoGe-2: Accurate monocular geometry with metric scale and sharp details.arXiv preprint arXiv:2507.02546,

    Ruicheng Wang, Sicheng Xu, Yue Dong, Yu Deng, Jianfeng Xiang, Zelong Lv, Guangzhong Sun, Xin Tong, and Jiaolong Yang. MoGe-2: Accurate monocular geometry with metric scale and sharp details.arXiv preprint arXiv:2507.02546,

    arXiv

  3. [1]

    Local-to-global panorama inpainting for locale-aware indoor lighting prediction.IEEE Trans

    Jiayang Bai, Zhen He, Shan Yang, Jie Guo, Zhenyu Chen, Yan Zhang, and Yanwen Guo. Local-to-global panorama inpainting for locale-aware indoor lighting prediction.IEEE Trans. Vis. Comput. Graph., 29(11):4405–4416, 2023. 2

    2023

  4. [2]

    Blender demo files

    Blender Foundation. Blender demo files. https://www. blender.org/download/demo-files/ , 2025. Ac- cessed: 2025-11-13. 5

    2025

  5. [3]

    GaSLight: Gaussian splats for spatially-varying lighting in HDR

    Christophe Bolduc, Yannick Hold-Geoffroy, Zhixin Shu, and Jean-Franc ¸ois Lalonde. GaSLight: Gaussian splats for spatially-varying lighting in HDR. InIEEE/CVF Int. Conf. Comput. Vis., 2025. 4

    2025

  6. [4]

    Intrinsic harmonization for illumination-aware compositing supplementary material

    Chris Careaga, S Mahdi H Miangoleh, and Ya ˘gız Aksoy. Intrinsic harmonization for illumination-aware compositing supplementary material. InACM SIGGRAPH Asia Conf.,

  7. [6]

    DiffusionLight-Turbo: Accelerated light probes for free via single-pass chrome ball inpainting

    Worameth Chinchuthakun, Pakkapon Phongthawee, Amit Raj, Varun Jampani, Pramook Khungurn, and Supasorn Suwa- janakorn. DiffusionLight-Turbo: Accelerated light probes for free via single-pass chrome ball inpainting. InArXiv, 2025. 2, 3

    2025

  8. [7]

    Blender Foundation, Stichting Blender Foundation, Amsterdam, 2025

    Blender Online Community.Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2025. 4

    2025

Show all 53 references
  1. [8]

    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 IEEE/CVF Int. Conf. Comput. Vis., 2023. 2

    2023

  2. [9]

    Rendering synthetic objects into real scenes: Bridging traditional and image-based graphics with global illumination and high dynamic range photography

    Paul Debevec. Rendering synthetic objects into real scenes: Bridging traditional and image-based graphics with global illumination and high dynamic range photography. InACM SIGGRAPH Conf., 1998. 1, 2

    1998

  3. [10]

    Estimating surface reflectance properties of a complex scene under captured natural illumination.ACM Trans

    Paul Debevec, Chris Tchou, Andrew Gardner, Tim Hawkins, Charis Poullis, Jessi Stumpfel, Andrew Jones, Nathaniel Yun, Per Einarsson, Therese Lundgren, Marcos Fajardo, and Philippe Martinez. Estimating surface reflectance properties of a complex scene under captured natural illu...

    2004

  4. [11]

    Debevec and Jitendra Malik.Recovering High Dy- namic Range Radiance Maps from Photographs

    Paul E. Debevec and Jitendra Malik.Recovering High Dy- namic Range Radiance Maps from Photographs. Association for Computing Machinery, New York, NY , USA, 1 edition,

  5. [12]

    BlenderKit: 3D assets, 2025

    Petr Dlouh´y, Vil´em Duha, Karol´ına H´userkov´a, Monika Ry- galov´a, Adam Krh ´anek, Eli ˇska Pant ˚uˇckov´a, Alex Hapon, Mike Radjabov, Amanpreet Bajwa, and Andreas Gajdo ˇs´ık. BlenderKit: 3D assets, 2025. 4

    2025

  6. [13]

    Deep neural models for illumination estimation and relighting: A survey.Comput

    Farshad Einabadi, Jean-Yves Guillemaut, and Adrian Hilton. Deep neural models for illumination estimation and relighting: A survey.Comput. Graph. Forum, 40(6):315–331, 2021. 2

    2021

  7. [14]

    Per- ception of lighting errors in image compositing

    James A Ferwerda, Jeremy Selan, and Fabio Pellacini. Per- ception of lighting errors in image compositing. InIS&T Color Imag. Conf., 2010. 1

    2010

  8. [15]

    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

  9. [16]

    Deep parametric indoor lighting estimation

    Marc-Andre Gardner, Yannick Hold-Geoffroy, Kalyan Sunkavalli, Christian Gagne, and Jean-Francois Lalonde. Deep parametric indoor lighting estimation. InIEEE/CVF Int. Conf. Comput. Vis., 2019. 2

    2019

  10. [17]

    Fast spatially-varying indoor lighting estimation

    Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-Francois Lalonde. Fast spatially-varying indoor lighting estimation. InIEEE/CVF Conf. Comput. Vis. Pattern Recog., 2019. 2, 5, 6, 3

    2019

  11. [18]

    Rrm: Relightable assets using radiance guided material ex- traction

    Diego Gomez, Julien Philip, Adrien Kaiser, and ´Elie Michel. Rrm: Relightable assets using radiance guided material ex- traction. InComp. Graph. Int. Conf., 2024. 5

    2024

  12. [19]

    Intrinsic image harmonization

    Zonghui Guo, Haiyong Zheng, Yufeng Jiang, Zhaorui Gu, and Bing Zheng. Intrinsic image harmonization. InIEEE/CVF Conf. Comput. Vis. Pattern Recog., 2021. 3

    2021

  13. [20]

    Deep sky modeling for single image out- door lighting estimation

    Yannick Hold-Geoffroy, Akshaya Athawale, and Jean- Franc ¸ois Lalonde. Deep sky modeling for single image out- door lighting estimation. InIEEE/CVF Conf. Comput. Vis. Pattern Recog., 2019. 2

    2019

  14. [21]

    Black Forest Labs. Flux. https://github.com/ black-forest-labs/flux, 2024. 5

    2024

  15. [22]

    Estimating the natural illumination conditions from a single outdoor image.Int

    Jean-Franc ¸ois Lalonde, Alexei A Efros, and Srinivasa G Narasimhan. Estimating the natural illumination conditions from a single outdoor image.Int. J. Comput. Vis., 98(2): 123–145, 2012. 2

    2012

  16. [23]

    Deep- Light: Learning illumination for unconstrained mobile mixed reality

    Chloe LeGendre, Wan-Chun Ma, Graham Fyffe, John Flynn, Laurent Charbonnel, Jay Busch, and Paul Debevec. Deep- Light: Learning illumination for unconstrained mobile mixed reality. InIEEE/CVF Conf. Comput. Vis. Pattern Recog.,

  17. [24]

    Learning illumination from diverse portraits

    Chloe LeGendre, Wan-Chun Ma, Rohit Pandey, Sean Fanello, Christoph Rhemann, Jason Dourgarian, Jay Busch, and Paul Debevec. Learning illumination from diverse portraits. In SIGGRAPH Asia 2020 Technical Communications, New York, NY , USA, 2020. Association for Computing Machinery. 2

    2020

  18. [25]

    Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image

    Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. InIEEE/CVF Conf. Comput. Vis. Pattern Recog., 2020. 2

    2020

  19. [26]

    Spatiotemporally consistent HDR indoor lighting estimation.ACM Trans

    Zhengqin Li, Li Yu, Mikhail Okunev, Manmohan Chandraker, and Zhao Dong. Spatiotemporally consistent HDR indoor lighting estimation.ACM Trans. Graph., 42(3):1–15, 2023. 2

    2023

  20. [27]

    DiffusionRenderer: Neural inverse and forward rendering with video diffusion models

    Ruofan Liang, Zan Gojcic, Huan Ling, Jacob Munkberg, Jon Hasselgren, Zhi-Hao Lin, Jun Gao, Alexander Keller, Nandita Vijaykumar, Sanja Fidler, and Zian Wang. DiffusionRenderer: Neural inverse and forward rendering with video diffusion models. InIEEE/CVF Conf. Comput. Vis. Patt...

  21. [28]

    LuxDiT: Light- ing estimation with video diffusion transformer

    Ruofan Liang, Kai He, Zan Gojcic, Igor Gilitschenski, Sanja Fidler, Nandita Vijaykumar, and Zian Wang. LuxDiT: Light- ing estimation with video diffusion transformer. InArXiv,

  22. [29]

    Lux post facto: Learning portrait performance relighting with conditional video diffusion and a hybrid dataset.IEEE/CVF Conf

    Yiqun Mei, Mingming He, Li Ma, Julien Philip, Wenqi Xian, David M George, Xueming Yu, Gabriel Dedic, Ahmet Levent Tas ¸el, Ning Yu, Vishal M Patel, and Paul Debevec. Lux post facto: Learning portrait performance relighting with conditional video diffusion and a hybrid dataset....

    2025

  23. [30]

    DiffusionLight: Light probes for free by painting a chrome ball

    Pakkapon Phongthawee, Worameth Chinchuthakun, Non- taphat Sinsunthithet, Amit Raj, Varun Jampani, Pramook Khungurn, and Supasorn Suwajanakorn. DiffusionLight: Light probes for free by painting a chrome ball. InIEEE/CVF Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 4, 5, 6

    2024

  24. [31]

    Infinigen indoors: Photorealistic indoor scenes us- ing procedural generation

    Alexander Raistrick, Lingjie Mei, Karhan Kayan, David Yan, Yiming Zuo, Beining Han, Hongyu Wen, Meenal Parakh, Stamatis Alexandropoulos, Lahav Lipson, Zeyu Ma, and Jia Deng. Infinigen indoors: Photorealistic indoor scenes us- ing procedural generation. InIEEE/CVF Conf. Comput....

    2024

  25. [32]

    Color transfer between images.IEEE Comp

    Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images.IEEE Comp. graph. appl., 21(5):34–41, 2002. 1

    2002

  26. [33]

    Morgan Kaufman, 2005

    Erik Reinhard, Greg Ward, Sumanta Pattanaik, and Paul De- bevec.High dynamic range imaging. Morgan Kaufman, 2005. 2

    2005

  27. [34]

    Illumination from shadows.IEEE Trans

    Imari Sato, Yoichi Sato, and Katsushi Ikeuchi. Illumination from shadows.IEEE Trans. Pattern Anal. Mach. Intell., 25 (3):290–300, 2003. 2

    2003

  28. [35]

    Neural illumination: Lighting prediction for indoor environments

    Shuran Song and Thomas Funkhouser. Neural illumination: Lighting prediction for indoor environments. InIEEE/CVF Conf. Comput. Vis. Pattern Recog., 2019. 2

    2019

  29. [36]

    Light- house: Predicting lighting volumes for spatially-coherent illu- mination

    Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snavely. Light- house: Predicting lighting volumes for spatially-coherent illu- mination. InIEEE/CVF Conf. Comput. Vis. Pattern Recog.,

  30. [37]

    Direct hdr capture of the sun and sky

    Jessi Stumpfel, Andrew Jones, Andreas Wenger, Chris Tchou, Tim Hawkins, and Paul Debevec. Direct hdr capture of the sun and sky. InACM SIGGRAPH Courses, 2006. 3

    2006

  31. [38]

    The perception of lighting inconsistencies in composite outdoor scenes.ACM Trans

    Minghui Tan, Jean-Franc ¸ois Lalonde, Lavanya Sharan, Holly Rushmeier, and Carol O’Sullivan. The perception of lighting inconsistencies in composite outdoor scenes.ACM Trans. Appl. Percept., 12(4), 2015. 1

    2015

  32. [39]

    Estimating spatially-varying lighting in urban scenes with disentangled representation

    Jiajun Tang, Yongjie Zhu, Haoyu Wang, Jun Hoong Chan, Si Li, and Boxin Shi. Estimating spatially-varying lighting in urban scenes with disentangled representation. InEur. Conf. Comput. Vis., 2022. 2

    2022

  33. [40]

    Raft: Recurrent all-pairs field transforms for optical flow

    Zachary Teed and Jia Den. Raft: Recurrent all-pairs field transforms for optical flow. InEur. Conf. Comput. Vis., 2020. 6

    2020

  34. [41]

    Spatiotem- porally consistent indoor lighting estimation with diffusion priors

    Mutian Tong, Rundi Wu, and Changxi Zheng. Spatiotem- porally consistent indoor lighting estimation with diffusion priors. InACM SIGGRAPH Conf., 2025. 2, 4, 5, 6, 7, 8

    2025

  35. [42]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianx- iao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jin- gren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, P...

    2025 arXiv

  36. [44]

    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. InIEEE/CVF Int. Conf. Comput. Vis., 2021. 2

    2021

  37. [45]

    RGB ↔X: Image decomposition and synthesis using material- and lighting-aware diffusion models

    Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloˇs Haˇsan. RGB ↔X: Image decomposition and synthesis using material- and lighting-aware diffusion models. InACM SIGGRAPH Conf., 2024. 3

    2024

  38. [46]

    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. InAssoc. Adv. of Art. Int., 2021. 2

    2021

  39. [47]

    All-weather deep outdoor lighting estimation

    Jinsong Zhang, Kalyan Sunkavalli, Yannick Hold-Geoffroy, Sunil Hadap, Jonathan Eisenman, and Jean-Franc ¸ois Lalonde. All-weather deep outdoor lighting estimation. InIEEE/CVF Conf. Comput. Vis. Pattern Recog., 2019. 2

    2019

  40. [48]

    ZeroComp: Zero- shot object compositing from image intrinsics via diffusion

    Zitian Zhang, Fr ´ed´eric Fortier-Chouinard, Mathieu Garon, Anand Bhattad, and Jean-Franc ¸ois Lalonde. ZeroComp: Zero- shot object compositing from image intrinsics via diffusion. InIEEE/CVF Winter Conf. App. Comput. Vis., 2025. 3

    2025

  41. [49]

    Irisformer: Dense vision transform- ers for single-image inverse rendering in indoor scenes

    Rui Zhu, Zhengqin Li, Janarbek Matai, Fatih Porikli, and Manmohan Chandraker. Irisformer: Dense vision transform- ers for single-image inverse rendering in indoor scenes. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., 2022. 2

    2022

  42. [50]

    Spatially- varying outdoor lighting estimation from intrinsics

    Yongjie Zhu, Yinda Zhang, Si Li, and Boxin Shi. Spatially- varying outdoor lighting estimation from intrinsics. In IEEE/CVF Conf. Comput. Vis. Pattern Recog., 2021. 2 Lighting in Motion: Spatiotemporal HDR Lighting Estimation Supplementary Material

    2021

  43. [51]

    For indoor scenes, we use the full indoor scenes provided by BlenderKit, gen- erating more cameras based on the original ones

    Data generation details We use Blender, paired with BlenderKit assets to procedu- rally generate indoor and outdoor renders. For indoor scenes, we use the full indoor scenes provided by BlenderKit, gen- erating more cameras based on the original ones. Since scenes are not alwa...

  44. [52]

    Gloss” (glossy) and “Mat

    HDRI map optimization details The predicted images from the network ˆI are cropped around the inpainted spheres. The same is done with the sphere mask, normals and position maps. The equirectan- gular HDRI is a Laplacian pyramid at a fixed resolution of 512x256 with 8 levels. ...

  45. [53]

    Additional results In complement to Tab. 2, Tab. 4 reports metrics on our se- quences test dataset for glossy and matte spheres. Sample predictions from The Laval Indoor Spatially Vary- ing HDR dataset [17] are presented in Fig. 7. More in-the-wild results are presented in Fig...

Pith tools

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