Pith. sign in

REVIEW 2 major objections 5 minor 45 references

Video diffusion models already know scene lighting; force them to paint a chrome ball and you get dynamic HDR environment maps from ordinary video.

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 · grok-4.5

2026-07-11 15:22 UTC pith:CDE4GB4Y

load-bearing objection Solid video extension of chrome-ball lighting estimation with real engineering value; the 'inherent' framing and photometric purity of the training data are the soft spots, not the core results. the 2 major comments →

arxiv 2607.04674 v1 pith:CDE4GB4Y submitted 2026-07-06 cs.CV

Video Generation Models Are Inherent Lighting Estimators

Ziqi Cai , Shuchen Weng , Kaiqi Liu , Zifeng Wang , Zhiquan Zhang , Minggui Teng , Han Jiang , Boxin Shi This is my paper
classification cs.CV
keywords video diffusion modelsdynamic lighting estimationHDR environment mapschrome ball inpaintingvideo inpaintingLoRA fine-tuninglight probes
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.

Recovering the full lighting of a casually filmed scene from one ordinary video is hard because the camera only ever sees a narrow, clipped view of the world. This paper claims that large video generation models already encode a workable understanding of physical lighting inside their generative priors. The authors unlock that knowledge by treating lighting recovery as a guided inpainting job: they mask the center of each frame and compel the model to fill it with a perfect mirrored chrome ball whose reflections must match the surrounding space and time. An HDR-aware autoencoder and light fine-tuning move the native low-dynamic-range model into the high-dynamic-range domain, while a hybrid training set of tonemapped panoramic videos and real static HDR panoramas supplies both motion context and absolute intensity anchors. The resulting environment maps stay coherent across frames and can light virtual objects with plausible highlights and shadows. If the claim holds, any future video diffusion backbone becomes a free lighting estimator for photorealistic compositing and relighting without extra sensors or multi-view capture.

Core claim

Modern video diffusion models are not merely synthesizers; they are inherently capable estimators of physical scene lighting. By reframing dynamic lighting estimation as the task of inpainting a synthetic chrome-ball light probe into a single in-the-wild video, the model can be made to output temporally coherent HDR environment maps that capture the scene's illumination.

What carries the argument

V-LITE: a chrome-ball inpainting pipeline that inserts a central masked sphere, conditions a LoRA-adapted flow-matching video transformer on the surrounding context plus a short text prompt, and unwraps the resulting HDR reflections into an environment-map sequence via an HDR-aware VAE with learnable tonemap adapters.

Load-bearing premise

That computationally inverse-tone-mapped panoramic videos, even after discarding more than half of them for photometric quality, together with a few hundred static real HDR images, give the model enough true linear intensity supervision to learn physical HDR lighting rather than merely plausible reflections.

What would settle it

On a held-out set of true multi-exposure HDR video captures that never appear in training, measure whether the recovered environment maps produce lower angular lighting error and higher temporal stability than the same model trained only on the paper's tonemapped data; a clear failure to match real intensities would refute the photometric claim.

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

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 / 5 minor

Summary. The paper claims that modern video diffusion models already encode usable physical lighting knowledge, and that this knowledge can be unlocked by reframing dynamic lighting estimation as chrome-ball video inpainting. V-LITE inserts a synthetic chrome ball into an in-the-wild LDR video, uses an HDR-aware VAE with learnable tonemap adapters plus LoRA fine-tuning of a Wan 2.1 backbone, and trains on a mixed set of VLM-filtered inverse-tonemapped panoramic videos and 800 real PolyHaven HDR images. The resulting temporally coherent HDR environment maps are evaluated on Editable Indoor and EnvMapNet metrics, ablations, qualitative insertion, and a small user study.

Significance. If the result holds, the work offers a practical route to dynamic HDR lighting from casual video without multi-view capture or specialized hardware, and it reframes video diffusion models as estimators rather than pure synthesizers. Strengths include a clear task reformulation inspired by VFX practice, an explicit HDR VAE design, a mixed-data training strategy with ablations that isolate the HDR VAE, LoRA, and static photometric anchors, competitive quantitative metrics against adapted image baselines, and demonstrated virtual-object insertion with a user study. The contribution is primarily empirical and systems-oriented rather than theoretical.

major comments (2)
  1. [Sec. 3 / Discussion 5.6 / Table 1] Sec. 3 and Discussion 5.6: the dynamic subset is built by applying three inverse tone-mapping operators to LDR PanoVid panoramas and discarding ~57% via VLM plausibility checks; the authors themselves note these videos are “not photometrically perfect.” Absolute linear intensities and high-frequency dynamics therefore rest largely on the 800 static PolyHaven images mixed at 10:1. Table 1 shows the video-only ablation nearly matches the full model on most metrics, so it remains unclear whether the system recovers true photometric HDR or mainly plausible reflections from generative priors. A load-bearing validation (e.g., intensity-ratio or exposure-bracket recovery on held-out real multi-exposure HDR video, or a controlled synthetic ground-truth comparison) is needed to support the “physical scene lighting” claim.
  2. [Sec. 5.1–5.2 / Table 1] Sec. 5.1–5.2 and Table 1: all primary baselines (DiffusionLight, DiffusionLight-Turbo, StyleLight) are single-image methods applied per-frame; DiffusionLight is further restricted to a 10-video subset because of cost. No comparison is reported against recent video-native lighting estimators (e.g., the spatiotemporal methods cited in Sec. 2.2 such as Li et al. 2023 or Tong et al. 2025). Without at least one video-aware baseline or a stronger temporal-consistency protocol, the claim of superior dynamic estimation is only partially supported.
minor comments (5)
  1. [Title / Abstract / Table 1] The “inherent” framing in the title and abstract is slightly overstated: the frozen-backbone ablation in Table 1 collapses (MSE 3.00), showing that LoRA and the HDR VAE are required. Soften the rhetoric to match the ablation evidence.
  2. [Sec. 4.2] Eqs. (3)–(4): the roles of the learnable 3D convolutions F/G versus the scalar scale/bias are not fully specified (initialization, whether they are shared across channels/time, training stability). A short clarification would help reproducibility.
  3. [Figs. 4, 6] Fig. 4 and Fig. 6 captions should state the tone-mapping used for visualization so readers can judge HDR fidelity versus display compression.
  4. [Sec. 5.5] User study (Sec. 5.5) reports percentages from 10 participants on 20 videos but omits inter-rater agreement, exact rating protocol, and statistical significance; a brief addition would strengthen the perceptual claim.
  5. [Sec. 4.3] Minor notation: M vs. M′ and the constant gray fill C for the mask are introduced without an explicit equation for the masked input construction; a single line would improve clarity.

Circularity Check

0 steps flagged

No circularity found: empirical adaptation of video diffusion priors for chrome-ball inpainting is validated on held-out photometric and rendering metrics without any result reducing to its inputs by construction.

full rationale

V-LITE is a standard empirical computer-vision method paper. It starts from a pretrained LDR video diffusion backbone (Wan 2.1), inserts learnable tonemap adapters into the VAE (Eqs. 3-4), freezes most weights and applies LoRA only to the diffusion Transformer, and trains the resulting model with a masked flow-matching objective (Eq. 7) on a hybrid dataset of inverse-tonemapped panoramic videos plus static real HDR images. Success is measured by external, previously published benchmarks (Editable Indoor MSE/SI-MSE/AER/LS and EnvMapNet AED/AS) plus qualitative virtual-object insertion; ablations (LDR baseline, frozen backbone, video-only) further isolate the contribution of each component. None of the reported numbers or qualitative claims is obtained by re-using a fitted parameter as a “prediction,” by defining a quantity in terms of itself, or by invoking an unverified uniqueness theorem from the same authors. Self-citations that appear (e.g., to the authors’ prior editing or panorama work) are ordinary related-work references and are not load-bearing for the derivation. The rhetorical framing that the models are “inherent” estimators is slightly overstated given that LoRA and the HDR VAE are required, yet this is marketing language, not a circular reduction. Dataset construction via inverse tonemapping plus VLM filtering raises a separate validity question about photometric fidelity, but that is outside the scope of circularity analysis. Consequently the derivation chain is self-contained and non-circular.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claim rests on the empirical premise that video diffusion priors already encode usable lighting, plus several engineering choices (mask placement, log-domain adapters, data mixture) that are free parameters or domain assumptions rather than derived necessities. No new physical entities are postulated; the chrome ball is a classical light probe.

free parameters (4)
  • LoRA rank r = 32
    Set to 32 for the diffusion Transformer attention layers; controls capacity of the adaptation.
  • mask weighting factor α = 5
    Emphasizes the chrome-ball region in the velocity loss (M' = 1+(α-1)M); chosen as 5.
  • dynamic:static mix ratio = 10:1
    Training samples mixed at 10:1 to balance temporal context against photometric anchors.
  • VAE two-stage step counts = 10k + 5k
    10k steps freezing VAE then 5k joint steps; schedule chosen for stable log-domain adaptation.
axioms (4)
  • domain assumption Pretrained video diffusion models already encode sufficient physical lighting priors that can be unlocked by chrome-ball inpainting.
    Stated in abstract and Sec. 1; supported only by the success of the adapted model, not independently proven.
  • domain assumption A central circular mask is irrelevant to global illumination and can be replaced by a synthetic chrome ball without changing the surrounding light field.
    Sec. 4.3; classical light-probe assumption transferred to video.
  • ad hoc to paper Inverse-tone-mapped LDR panoramas filtered by a VLM retain usable HDR intensity ratios for supervision.
    Sec. 3; 57% of sequences discarded, remaining treated as training targets despite acknowledged photometric imperfections.
  • domain assumption Flow-matching latent diffusion with LoRA preserves generative spatio-temporal priors while aligning to HDR latents.
    Sec. 4.2–4.3; standard transfer-learning assumption.
invented entities (2)
  • HDR-aware VAE with learnable tonemap adapters F/G no independent evidence
    purpose: Map linear HDR into the log-domain latent space of an LDR-pretrained VAE and back.
    New architectural component introduced in Sec. 4.2; no independent existence outside this pipeline.
  • V-LITESet no independent evidence
    purpose: Provide mixed dynamic tonemapped videos and static real HDR images for co-training.
    Constructed dataset (Sec. 3); not previously published as a lighting-estimation benchmark.

pith-pipeline@v1.1.0-grok45 · 17469 in / 2891 out tokens · 31367 ms · 2026-07-11T15:22:22.440328+00:00 · methodology

0 comments
read the original abstract

Recovering dynamic environment maps from a single in-the-wild video is crucial for photorealistic rendering, yet remains a challenge. Recent video generation models can produce photorealistic scenes with complex lighting, possessing an inherent understanding of lighting. In this paper, we introduce V-LITE (Video generation models are inherent lighting estimators), a framework that unlocks this internal knowledge by reframing lighting estimation as a guided video inpainting task. Inspired by VFX industry practices, we insert a synthetic chrome ball into the scene to compel the model to generate physically plausible reflections from the surrounding spatio-temporal context. To bridge the gap from LDR-native models to the HDR domain, we design an HDR-aware VAE and employ an efficient LoRA-based fine-tuning strategy. We then construct a mixed dataset comprising high-fidelity HDR images to provide realistic HDR priors, and in-the-wild HDR videos to provide dynamic spatio-temporal context. Extensive experiments demonstrate that V-LITE produces temporally coherent HDR environment maps, revealing that modern video diffusion models are not merely synthesizers but also powerful, inherently capable estimators of physical scene lighting.

Figures

Figures reproduced from arXiv: 2607.04674 by Boxin Shi, Han Jiang, Kaiqi Liu, Minggui Teng, Shuchen Weng, Zhiquan Zhang, Zifeng Wang, Ziqi Cai.

Figure 1
Figure 1. Figure 1: An illustration of our V-LITE framework. (a) Inspired by on-set VFX practices, we reformulate dynamic lighting estimation as the task of inserting a virtual light probe into any in-the-wild LDR video. (b) We implement this concept as a video inpainting task, compelling the video diffusion model to leverage its inherent priors and render a physically plausible chrome ball that captures the dynamic illuminat… view at source ↗
Figure 2
Figure 2. Figure 2: Samples from the V-LITESet dataset. Each row corresponds to a video, and the images in the row represent different timestamps sampled from the video. physics-grounded priors by jointly reasoning about geometry, BRDFs, and il￾lumination from a single image [31]. For outdoor scenes, single-image methods often predict parametric sun/sky lighting [18]. Despite these advances, single￾image approaches typically … view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the V-LITE pipeline. (a) HDR VAE processing. To operate in log-domain space, an input HDR video xHDR is mapped by a learnable tone mapping network F to xlog, encoded by E, and decoded by DHDR into x ′ log. An inverse net￾work G then reconstructs the HDR video x ′ HDR. We pretrain F and G before jointly optimizing them with DHDR. (b) Inherent lighting estimation. For an LDR video X masked by M (… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison with baseline lighting estimation models. We compare our method against three methods: StyleLight [37] (SL), DiffusionLight [29] (DL), and DiffusionLightTurbo [11] (DLT). For each block, the first row shows input video frames, followed by the predicted lighting conditions from each method, and ground truth (GT) in the final row. Our method produces lighting estimates that closely mat… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results on in-the-wild videos. For visualization purposes, all frames are tone-mapped from their original HDR format. ing solely on tonemapped HDR video data limits the model’s ability to predict accurate physical HDR values. Instead, our full model incorporates both the HDR VAE and the LoRA-based fine-tuning, utilizing a mixed-data training strategy. It outperforms all ablation variants, demon… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of virtual object insertion. We compare our method against StyleLight [37], DiffusionLight [29], and DiffusionLightTurbo [11]. To ensure a fair comparison, we use the dynamic HDR environment map sequence estimated by each method to illuminate a virtual object within a fixed 3D scene. The first column displays the input video frame, followed by the rendered results from each competing… view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of zero-shot generalization by varying probe position and size [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] 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

45 extracted references · 6 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2502.13923 (2025)

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-VL technical report. arXiv preprint arXiv:2502.13923 (2025)

    Pith/arXiv arXiv 2025

  2. [2]

    In: International Conference on Computer Vision (2021)

    Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: International Conference on Computer Vision (2021)

    2021

  3. [3]

    Blender Online Community: Blender.https://www.blender.org, accessed: 2026- 03-01

    2026

  4. [4]

    In: International Conference on Computer Vision (2025)

    Bolduc, C., Hold-Geoffroy, Y., Shu, Z., Lalonde, J.F.: GaSLight: Gaussian splats for spatially-varying lighting in hdr. In: International Conference on Computer Vision (2025)

    2025

  5. [5]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    Cai, Z., Jiang, K., Chen, S.Y., Lai, Y.K., Fu, H., Shi, B., Gao, L.: Real-time 3D- aware portrait video relighting. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    2024

  6. [6]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2025)

    Cai, Z., Weng, S., Xia, Y., Shi, B.: PhyS-EdiT: Physics-aware semantic image editing with text description. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2025)

    2025

  7. [7]

    arXiv preprint arXiv:2604.07966 (2026)

    Cai, Z., Yang, T., Chang, Z., Li, S., Jiang, H., Weng, S., Shi, B.: Lighting- grounded video generation with renderer-based agent reasoning. arXiv preprint arXiv:2604.07966 (2026)

    Pith/arXiv arXiv 2026

  8. [8]

    In: Advances in Neural Information Processing Systems (2023)

    Chang, Z., Weng, S., Zhang, P., Li, Y., Li, S., Shi, B.: L-CAD: Language-based col- orization with any-level descriptions using diffusion priors. In: Advances in Neural Information Processing Systems (2023)

    2023

  9. [9]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (2025)

    Chen, K., Ramanan, D., Khurana, T.: Using diffusion priors for video amodal segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (2025)

    2025

  10. [10]

    In: International Conference on Computer Vision (2023)

    Chen, S., Sun, P., Song, Y., Luo, P.: DiffusionDet: Diffusion model for object detection. In: International Conference on Computer Vision (2023)

    2023

  11. [11]

    IEEE Transactions on Pattern Analysis and Machine In- telligence (2026)

    Chinchuthakun, W., Phongthawee, P., Raj, A., Jampani, V., Khungurn, P., Suwa- janakorn, S.: DiffusionLight-Turbo: Accelerated light probes for free via single-pass chrome ball inpainting. IEEE Transactions on Pattern Analysis and Machine In- telligence (2026)

    2026

  12. [12]

    In: ACM SIGGRAPH Conference Papers (2008)

    Debevec, P.: Rendering synthetic objects into real scenes: Bridging traditional and image-based graphics with global illumination and high dynamic range photogra- phy. In: ACM SIGGRAPH Conference Papers (2008)

    2008

  13. [13]

    In: International Conference on Computer Vision (2023)

    Feng, R., Gao, Y., Tse, T.H.E., Ma, X., Chang, H.J.: DiffPose: Spatiotemporal dif- fusion model for video-based human pose estimation. In: International Conference on Computer Vision (2023)

    2023

  14. [14]

    arXiv preprint arXiv:2506.09113 (2025)

    Gao, Y., Guo, H., Hoang, T., Huang, W., Jiang, L., Kong, F., Li, H., Li, J., Li, L., Li, X., et al.: Seedance 1.0: Exploring the boundaries of video generation models. arXiv preprint arXiv:2506.09113 (2025)

    Pith/arXiv arXiv 2025

  15. [15]

    ACM Transactions on Graphics (2017)

    Gardner, M.A., Sunkavalli, K., Yumer, E., Shen, X., Gambaretto, E., Gagné, C., Lalonde, J.F.: Learning to predict indoor illumination from a single image. ACM Transactions on Graphics (2017)

    2017

  16. [16]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019) Video Generation Models Are Inherent Lighting Estimators 17

    Garon,M.,Sunkavalli,K.,Hadap,S.,Carr,N.,Lalonde,J.F.:Fastspatially-varying indoor lighting estimation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019) Video Generation Models Are Inherent Lighting Estimators 17

    2019

  17. [17]

    Haven, P.: Poly Haven: The public 3d asset library.https://polyhaven.com/, accessed: 2026-03-01

    2026

  18. [18]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017)

    Hold-Geoffroy, Y., Sunkavalli, K., Hadap, S., Gambaretto, E., Lalonde, J.F.: Deep outdoor illumination estimation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017)

    2017

  19. [19]

    In: International Conference on Learning Representations (2023)

    Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: CogVideo: Large-scale pretrain- ing for text-to-video generation via transformers. In: International Conference on Learning Representations (2023)

    2023

  20. [20]

    In: International Conference on Learning Representations (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2022)

    2022

  21. [21]

    In: International Conference on Learning Representations (2025)

    Huang, T., Yan, Z., Zhao, Y., Lee, G.H.: ComPC: Completing a 3D point cloud with 2D diffusion priors. In: International Conference on Learning Representations (2025)

    2025

  22. [22]

    In: International Conference on Computer Vision (2025)

    Jiang, Z., Han, Z., Mao, C., Zhang, J., Pan, Y., Liu, Y.: VACE: All-in-one video creation and editing. In: International Conference on Computer Vision (2025)

    2025

  23. [23]

    In: Association for the Advancement of Artificial Intelligence (2025)

    Lan, Y., Cui, Z., Liu, C., Peng, J., Wang, N., Luo, X., Liu, D.: Exploiting diffusion prior for real-world image dehazing with unpaired training. In: Association for the Advancement of Artificial Intelligence (2025)

    2025

  24. [24]

    ACM Transactions on Graphics (2023)

    Li, Z., Yu, L., Okunev, M., Chandraker, M., Dong, Z.: Spatiotemporally consistent HDR indoor lighting estimation. ACM Transactions on Graphics (2023)

    2023

  25. [25]

    In: European Conference on Computer Vision (2024)

    Liang, R., Gojcic, Z., Nimier-David, M., Acuna, D., Vijaykumar, N., Fidler, S., Wang, Z.: Photorealistic object insertion with diffusion-guided inverse rendering. In: European Conference on Computer Vision (2024)

    2024

  26. [26]

    In: Advances in Neural Information Processing Systems (2025)

    Liang, R., He, K., Gojcic, Z., Gilitschenski, I., Fidler, S., Vijaykumar, N., Wang, Z.: LuxDiT: Lighting estimation with video diffusion transformer. In: Advances in Neural Information Processing Systems (2025)

    2025

  27. [27]

    In: International Conference on Learning Representations (2023)

    Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: International Conference on Learning Representations (2023)

    2023

  28. [28]

    In: International Conference on Computer Vision (2019)

    Murmann, L., Gharbi, M., Aittala, M., Durand, F.: A dataset of multi-illumination images in the wild. In: International Conference on Computer Vision (2019)

    2019

  29. [29]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    Phongthawee, P., Chinchuthakun, W., Sinsunthithet, N., Raj, A., Jampani, V., Khungurn, P., Suwajanakorn, S.: DiffusionLight: Light probes for free by paint- ing a chrome ball. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024)

    2024

  30. [30]

    In: ACM SIGGRAPH Conference Papers (2024)

    Rao, P., Fox, G., Meka, A., BR, M., Zhan, F., Weyrich, T., Bickel, B., Pfister, H., Matusik, W., Elgharib, M., et al.: Lite2Relight: 3D-aware single image portrait relighting. In: ACM SIGGRAPH Conference Papers (2024)

    2024

  31. [31]

    In: International Conference on Computer Vision (2019)

    Sengupta, S., Gu, J., Kim, K., Liu, G., Jacobs, D.W., Kautz, J.: Neural inverse rendering of an indoor scene from a single image. In: International Conference on Computer Vision (2019)

    2019

  32. [32]

    Song, S., Funkhouser, T.: Neural illumination: Lighting prediction for indoor envi- ronments.In:IEEE/CVFConferenceonComputerVisionandPatternRecognition (2019)

    2019

  33. [33]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)

    Srinivasan, P.P., Mildenhall, B., Tancik, M., Barron, J.T., Tucker, R., Snavely, N.: Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)

    2020

  34. [34]

    In: Advances in Neural Informa- tion Processing Systems (2023) 18 Z

    Tang, J., Zhong, H., Weng, S., Shi, B.: LuminAIRe: Illumination-aware conditional image repainting for lighting-realistic generation. In: Advances in Neural Informa- tion Processing Systems (2023) 18 Z. Cai et al

    2023

  35. [35]

    arXiv preprint arXiv:2508.08384 (2025)

    Tong, M., Wu, R., Zheng, C.: Spatiotemporally consistent indoor lighting estima- tion with diffusion priors. arXiv preprint arXiv:2508.08384 (2025)

    Pith/arXiv arXiv 2025

  36. [36]

    arXiv preprint arXiv:2503.20314 (2025)

    Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.W., Chen, D., Yu, F., Zhao, H., Yang, J., et al.: Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314 (2025)

    Pith/arXiv arXiv 2025

  37. [37]

    In: European Conference on Computer Vision (2022)

    Wang, G., Yang, Y., Loy, C.C., Liu, Z.: StyleLight: HDR panorama generation for lighting estimation and editing. In: European Conference on Computer Vision (2022)

    2022

  38. [38]

    Wang, Q., Zhao, Y., Ma, J., Li, J.: How to use diffusion priors under sparse views? In: Advances in Neural Information Processing Systems (2024)

    2024

  39. [39]

    In: IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (2024)

    Wang, X., Xiao, S., Liang, X.: LightOctree: Lightweight 3D spatially-coherent in- door lighting estimation. In: IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (2024)

    2024

  40. [40]

    In: International Conference on Learning Representations (2024)

    Wang, Y., He, Y., Li, Y., Li, K., Yu, J., Ma, X., Li, X., Chen, G., Chen, X., Wang, Y., et al.: InternVid: A large-scale video-text dataset for multimodal understanding and generation. In: International Conference on Learning Representations (2024)

    2024

  41. [41]

    arXiv preprint arXiv:2509.20328 (2025)

    Wiedemer, T., Li, Y., Vicol, P., Gu, S.S., Matarese, N., Swersky, K., Kim, B., Jaini, P., Geirhos, R.: Video models are zero-shot learners and reasoners. arXiv preprint arXiv:2509.20328 (2025)

    Pith/arXiv arXiv 2025

  42. [42]

    In: Advances in Neural Information Pro- cessing Systems (2025)

    Xia, Y., Weng, S., Yang, S., Liu, J., Zhu, C., Teng, M., Jia, Z., Jiang, H., Shi, B.: PanoWan: Lifting diffusion video generation models to 360°with latitude/longitude-aware mechanisms. In: Advances in Neural Information Pro- cessing Systems (2025)

    2025

  43. [43]

    In: European Conference on Computer Vision (2024)

    Yang, T., Wu, R., Ren, P., Xie, X., Zhang, L.: Pixel-aware stable diffusion for real- istic image super-resolution and personalized stylization. In: European Conference on Computer Vision (2024)

    2024

  44. [44]

    In: International Conference on Learning Representations (2025)

    Yang, Z., Teng, J., Zheng, W., Ding, M., Huang, S., Xu, J., Yang, Y., Hong, W., Zhang, X., Feng, G., et al.: CogVideoX: Text-to-video diffusion models with an expert transformer. In: International Conference on Learning Representations (2025)

    2025

  45. [45]

    In: Association for the Advancement of Artificial Intelligence (2021)

    Zhan, F., Zhang, C., Yu, Y., Chang, Y., Lu, S., Ma, F., Xie, X.: EMLight: Light- ing estimation via spherical distribution approximation. In: Association for the Advancement of Artificial Intelligence (2021)

    2021