pith. sign in

arxiv: 2606.22718 · v1 · pith:TNHWA4P7new · submitted 2026-06-21 · 💻 cs.CV

Generative Relightable Avatars

Kunwar Maheep Singh , Christian Theobalt , Rishabh Dabral This is my paper

Pith reviewed 2026-06-26 10:23 UTC · model grok-4.3

classification 💻 cs.CV
keywords relightable avatarsfull-body humansenvironment map relightingdiffusion modelsphysics-based renderinggenerative refinementperson-specific modelsvideo-to-video translation
0
0 comments X

The pith

A hybrid pipeline of physics-based rendering and generative refinement produces higher-quality relightable full-body human avatars than prior methods.

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

The paper introduces Generative Relightable Avatars as a person-specific technique for free-view rendering and relighting of full-body humans under arbitrary environment maps. It starts from the premise that fine appearance details form a one-to-many mapping best addressed by combining deterministic physics with probabilistic generation rather than pure regression. The method optimizes materials on a tracked mesh to create a controllable coarse rendering, refines textures via a feed-forward network for pose and illumination effects, and applies a video diffusion model to reach high-detail output while retaining lighting control. Experiments show this yields better perceptual results than earlier relightable avatar approaches. The work focuses on maintaining 3D consistency and temporal coherence across novel views and lights.

Core claim

GRA follows a hybrid approach that combines controllable, physics-grounded relighting with probabilistic refinement. Starting from a tracked animated mesh, material parameters are optimized in UV-space and rendered as a coarse relit appearance under a target HDR environment map; a feed-forward model then captures pose-dependent texture dynamics and illumination effects beyond simplified reflectance; finally a fine-tuned video-to-video diffusion model transforms the renderings into temporally coherent high-detail videos while preserving 3D control, using an error-recycling strategy for long sequences.

What carries the argument

Three-stage hybrid pipeline of UV-space material optimization with physics rendering, feed-forward texture refinement, and error-recycling video-to-video diffusion.

Load-bearing premise

The tracked animated mesh is accurate enough and the coarse physics render stable enough that later neural stages can add detail without breaking 3D consistency or lighting control.

What would settle it

Apply the full pipeline to input meshes with known tracking inaccuracies and measure whether output videos lose multi-view consistency or fail to match new environment maps under controlled relighting tests.

Figures

Figures reproduced from arXiv: 2606.22718 by Christian Theobalt, Kunwar Maheep Singh, Rishabh Dabral.

Figure 1
Figure 1. Figure 1: We present Generative Relightable Avatars, a novel method that enables pho￾torealistic relighting of full-body humans in motion. From pose, viewpoint, and illumi￾nation as input, our method generates temporally consistent, physically grounded ren￾derings under novel lighting, which also qualitatively generalizes to out-of-distribution one-light-at-a-time (OLAT) setups and near-field lighting effects. This … view at source ↗
Figure 2
Figure 2. Figure 2: Overview. Given skeletal motion θ, viewpoint κ, and environment illumina￾tion E, GRA outputs a photorealistic rendering of the character avatar in three steps: (a) A person-specific animatable mesh predicts pose-dependent geometry M, relit un￾der a microfacet BRDF to produce a coarse appearance (Sec. 3.1), which is refined by (b) RelightNet to capture illumination- and pose-dependent effects (Sec. 3.2). (c… view at source ↗
Figure 2
Figure 2. Figure 2: This approximation has two key limitations: (i) due to the restricted [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Generative Refinement. Given the coarse renderings of the character from a target viewpoint, we finetune the WAN2.1-VACE architecture to produce fine￾grained renderings. Environment illumination is introduced through Atemporal Cross￾Attention that performs per-frame attention between Envmap tokens and video tokens. backbone based on Latent Flow Matching (LFM) [21], adapted to incorporate lighting controls.… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative Results. GRA generates photorealistic free-view renderings of dynamic humans under novel poses, viewpoints, and illumination. Characters are seam￾lessly placed into diverse real-world environments, demonstrating realistic relighting, coherent shading, and consistent appearance under challenging lighting conditions. No￾tably, our physics-grounded model is able to generalize to OLAT lighting cond… view at source ↗
Figure 6
Figure 6. Figure 6: Physics Grounded Applications. We demonstrate two capabilities of our w/o RelightNet model enabled by its explicit physics-grounded design. (Left) A local near-field point light source produces physically plausible shading, including arm self￾shadowing. For comparison, our full model approximates the same light as far-field illumination by projecting it onto the lightstage. (Right) Relightable texture edit… view at source ↗
Figure 5
Figure 5. Figure 5: Starting from a Subject 4 check￾point, we finetune our video model on 170 frames of Subject 1. The adapted model produces photorealistic, temporally consis￾tent relighting for the new identity despite minimal supervision. by our pipeline raises a natural question about generalization across identities. Our few-shot adaptation experiment ( [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative Comparison. We compare our method with state-of-the-art re￾lightable avatar methods: Relighting4D (R4D) [3] augmented with a DDC-tracked template, IntrinsicAvatar [37], MeshAvatar [2], and DNF-Avatar [13]. Our approach consistently improves visual realism and relighting quality. appearance with video generative models allows us to have 3D control as well as photorealism at the same time. Intere… view at source ↗
Figure 8
Figure 8. Figure 8: Ablation of key design components. The few-shot adaptation variant, while perceptually realistic, exhibits imprecise shading, while w/o Atemporal Lighting Conditioning lacks global illumination context, resulting in inaccurate shading. lighting from a separate conditioning signal. Finetuning overcomes this bias by teaching the model to treat the coarse renderings as a shading target [PITH_FULL_IMAGE:figur… view at source ↗
Figure 9
Figure 9. Figure 9: Metrics over clip length, w/ and w/o Error Recycling. clips, generated every 1000 testing frames ( [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Ablation of key design components. Microfacet rendering exhibits ar￾tifacts due to simplified light transport assumptions. RelightNet improves global il￾lumination but smooths high-frequency details due to pose-to-appearance ambiguity and mesh-based representation. Without RelightNet initialization, generative refine￾ment produces incorrect shading. Without VACE finetuning, appearance degrades and lightin… view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative ablation of tracking robustness. – 0:49–1:07 — Generalization to out-of-distribution OLAT lighting condi￾tions, highlighting self-shadowing effects. – 1:09–1:21 — Qualitative evaluation of a model variant enabling near-field relighting without retraining, contrasted with far-field relighting. Occlusion effects of the light source by the subject’s arm are also illustrated. – 1:22–1:28 — Relight… view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative comparison against baseline methods in the testing sequence of additional two subjects. error recycling for long-horizon generation can introduce minor temporal color jitter, as the model may imperfectly disentangle shading variations from injected error during continuation. This artifact could be mitigated by incorporating a lightweight post-processing deflickering module, such as blind video… view at source ↗
read the original abstract

We present Generative Relightable Avatars (GRA), a person-specific method for photorealistic free-view rendering and environment-map relighting of full-body humans. We postulate that modeling fine-grained appearance details is inherently a one-to-many problem that can benefit from a generative formulation. In contrast to fully regressive relightable avatar methods, GRA follows a hybrid approach that combines controllable, physics-grounded relighting with probabilistic refinement. Starting from a tracked animated mesh, we optimize material parameters in UV-space and render a coarse relit appearance under a target HDR environment map. Next, we refine the textures with a feed-forward model to capture pose-dependent texture dynamics and illumination effects beyond simplified reflectance assumptions. Finally, a fine-tuned video-to-video diffusion model transforms the physically grounded renderings into temporally coherent, high-detail videos while preserving 3D control, with an error-recycling strategy for generating long videos. Experimental evaluations demonstrate our method's improved perceptual quality over prior relightable avatar baselines. Project Page: https://vcai.mpi-inf.mpg.de/projects/GRA/

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper presents Generative Relightable Avatars (GRA), a person-specific hybrid method for photorealistic free-view rendering and environment-map relighting of full-body humans. Starting from a tracked animated mesh, it optimizes UV-space material parameters to produce a coarse physics-based render under a target HDR environment map, refines the result with a feed-forward model to capture pose-dependent texture dynamics, and applies a fine-tuned video-to-video diffusion model with an error-recycling strategy to generate temporally coherent high-detail output while preserving 3D control and lighting fidelity. The central claim is improved perceptual quality relative to prior relightable avatar baselines.

Significance. If the hybrid pipeline maintains 3D consistency and lighting control while adding generative detail, the work would offer a practical advance over purely regressive relightable avatar methods by explicitly addressing the one-to-many character of fine appearance. The error-recycling mechanism for long-video generation is a concrete engineering contribution that could see adoption in animation pipelines if shown to scale without drift.

major comments (2)
  1. [Abstract] Abstract: the assertion of 'improved perceptual quality' is presented without any quantitative metrics, user-study results, ablation tables, or dataset statistics; because this is the central empirical claim, the experiments section must supply these (including error bars and baseline comparisons) for the claim to be evaluable.
  2. [Method description (implied in abstract)] The weakest assumption identified in the method description—that the input tracked mesh and coarse physics render remain sufficiently stable for the feed-forward and diffusion stages to add detail without introducing 3D-inconsistent artifacts or breaking lighting control—is load-bearing; the manuscript must demonstrate this stability (e.g., via controlled ablation on mesh noise or render fidelity) in §4 or §5.
minor comments (1)
  1. [Abstract] The manuscript should clarify the exact architecture and training objective of the feed-forward refinement network and the conditioning signals used for the video diffusion model.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive review. We address each major comment below and commit to revisions that strengthen the empirical support and validation of key assumptions in the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion of 'improved perceptual quality' is presented without any quantitative metrics, user-study results, ablation tables, or dataset statistics; because this is the central empirical claim, the experiments section must supply these (including error bars and baseline comparisons) for the claim to be evaluable.

    Authors: We agree that the central claim of improved perceptual quality requires explicit quantitative and statistical support to be fully evaluable. While the current experiments section provides qualitative comparisons and perceptual evaluations against baselines, we will revise it to include quantitative metrics (LPIPS, FID) with error bars, a user study with statistical analysis, ablation tables, and dataset statistics. These additions will directly support the abstract claim. revision: yes

  2. Referee: [Method description (implied in abstract)] The weakest assumption identified in the method description—that the input tracked mesh and coarse physics render remain sufficiently stable for the feed-forward and diffusion stages to add detail without introducing 3D-inconsistent artifacts or breaking lighting control—is load-bearing; the manuscript must demonstrate this stability (e.g., via controlled ablation on mesh noise or render fidelity) in §4 or §5.

    Authors: We concur that explicit validation of the stability assumption is necessary. In the revised manuscript we will add a controlled ablation study (in §4 or §5) that introduces controlled mesh noise and render-fidelity variations, then measures resulting 3D consistency and lighting fidelity in the final output. This will quantify the robustness of the hybrid pipeline. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper describes a hybrid pipeline of UV-space material optimization, coarse physics-based rendering under HDR maps, feed-forward texture refinement for pose-dependent effects, and a fine-tuned video diffusion model with error recycling. No equations, derivations, or self-referential definitions appear in the provided text that reduce any claimed output (e.g., relit appearance or perceptual quality) to fitted inputs by construction. The method assembles standard rendering and generative components without load-bearing self-citations, uniqueness theorems imported from prior author work, or renaming of known results as novel derivations. The central claims rest on the independent engineering integration of these stages rather than tautological reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Review performed on abstract only; full paper text unavailable, so ledger is necessarily incomplete. The method explicitly starts from a tracked mesh and assumes standard reflectance and diffusion components.

axioms (2)
  • domain assumption A tracked animated mesh of the subject is available as input
    Abstract states 'Starting from a tracked animated mesh'
  • domain assumption Standard physically based rendering equations are sufficient for the coarse stage
    Abstract invokes 'physics-grounded relighting' and 'simplified reflectance assumptions'

pith-pipeline@v0.9.1-grok · 5717 in / 1377 out tokens · 35275 ms · 2026-06-26T10:23:48.990524+00:00 · methodology

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

52 extracted references · 1 canonical work pages

  1. [1]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

    Chaturvedi, S., Ren, M., Hold-Geoffroy, Y., Liu, J., Dorsey, J., Shu, Z.: Synthlight: Portrait relighting with diffusion model by learning to re-render synthetic faces. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

    2025

  2. [2]

    arXiv preprint arXiv:2407.08414 (2024)

    Chen, Y., Zheng, Z., Li, Z., Xu, C., Liu, Y.: Meshavatar: Learning high-quality triangular human avatars from multi-view videos. arXiv preprint arXiv:2407.08414 (2024)

    arXiv 2024

  3. [3]

    In: Euro- pean Conference on Computer Vision

    Chen, Z., Liu, Z.: Relighting4d: Neural relightable human from videos. In: Euro- pean Conference on Computer Vision. pp. 606–623. Springer (2022)

    2022

  4. [4]

    ACM Trans

    Cook, R.L., Torrance, K.E.: A reflectance model for computer graphics. ACM Trans. Graph.1(1), 7–24 (Jan 1982).https://doi.org/10.1145/357290.357293, https://doi.org/10.1145/357290.357293

    work page doi:10.1145/357290.357293 1982

  5. [5]

    Bermano, A., Theobalt, C.: Practilight: Practical light control using foundational diffusion models

    Erel, Y., Dabral, R., Golyanik, V., H. Bermano, A., Theobalt, C.: Practilight: Practical light control using foundational diffusion models. ACM Transactions on Graphics (TOG)44(6), 1–11 (2025)

    2025

  6. [6]

    arXiv preprint arXiv:1704.00090 (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. arXiv preprint arXiv:1704.00090 (2017)

    Pith/arXiv arXiv 2017

  7. [7]

    ACM Transactions on Graphics (ToG)40(4), 1–16 (2021)

    Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. ACM Transactions on Graphics (ToG)40(4), 1–16 (2021)

    2021

  8. [8]

    In: European Conference on Computer Vision (ECCV) (2010)

    He, K., Sun, J., Tang, X.: Guided image filtering. In: European Conference on Computer Vision (ECCV) (2010)

    2010

  9. [9]

    In: SIGGRAPH Asia 2024 Conference Papers

    He, M., Clausen, P., Taşel, A.L., Ma, L., Pilarski, O., Xian, W., Rikker, L., Yu, X., Burgert,R.,Yu,N.,etal.:Diffrelight:Diffusion-basedfacialperformancerelighting. In: SIGGRAPH Asia 2024 Conference Papers. pp. 1–12 (2024)

    2024

  10. [10]

    arXiv preprint arXiv:1503.02531 (2015)

    Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)

    Pith/arXiv arXiv 2015

  11. [11]

    Hu, E., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Chen, W.: Lora: Low-rank adaptation of large language models (2021)

    2021

  12. [12]

    In: ICCV (2023)

    Iqbal, U., Caliskan, A., Nagano, K., Khamis, S., Molchanov, P., Kautz, J.: Rana: Relightable articulated neural avatars. In: ICCV (2023)

    2023

  13. [13]

    In: Proceedings of the IEEE/CVF International Con- ference on Computer Vision

    Jiang, Z., Wang, S., Tang, S.: Dnf-avatar: Distilling neural fields for real-time an- imatable avatar relighting. In: Proceedings of the IEEE/CVF International Con- ference on Computer Vision. pp. 6383–6394 (2025)

    2025

  14. [14]

    arXiv preprint arXiv:2503.07598 (2025)

    Jiang, Z., Han, Z., Mao, C., Zhang, J., Pan, Y., Liu, Y.: Vace: All-in-one video creation and editing. arXiv preprint arXiv:2503.07598 (2025)

    Pith/arXiv arXiv 2025

  15. [15]

    arXiv preprint arXiv:2406.07520 (2024)

    Jin, H., Li, Y., Luan, F., Xiangli, Y., Bi, S., Zhang, K., Xu, Z., Sun, J., Snavely, N.: Neural gaffer: Relighting any object via diffusion. arXiv preprint arXiv:2406.07520 (2024)

    arXiv 2024

  16. [16]

    Kajiya,J.T.:Therenderingequation.In:Proceedingsofthe13thannualconference on Computer graphics and interactive techniques. pp. 143–150 (1986)

    1986

  17. [17]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Kim, H., Jang, M., Yoon, W., Lee, J., Na, D., Woo, S.: Switchlight: Co-design of physics-driven architecture and pre-training framework for human portrait re- lighting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25096–25106 (2024)

    2024

  18. [18]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Lei, C., Ren, X., Zhang, Z., Chen, Q.: Blind video deflickering by neural filtering with a flawed atlas. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10439–10448 (2023) 16 K.M. Singh et al

    2023

  19. [19]

    arXiv preprint arXiv:2510.09212 (2025)

    Li, W., Pan, W., Luan, P.C., Gao, Y., Alahi, A.: Stable video infinity: Infinite- length video generation with error recycling. arXiv preprint arXiv:2510.09212 (2025)

    arXiv 2025

  20. [20]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

    Liang, R., Gojcic, Z., Ling, H., Munkberg, J., Hasselgren, J., Lin, C.H., Gao, J., Keller, A., Vijaykumar, N., Fidler, S., Wang, Z.: Diffusionrenderer: Neural in- verse and forward rendering with video diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

    2025

  21. [21]

    arXiv preprint arXiv:2210.02747 (2022)

    Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. arXiv preprint arXiv:2210.02747 (2022)

    Pith/arXiv arXiv 2022

  22. [22]

    arXiv preprint arXiv:2407.16124 (2024)

    Liu, J., Qu, Y., Yan, Q., Zeng, X., Wang, L., Liao, R.: Fr\’echet video motion distance: A metric for evaluating motion consistency in videos. arXiv preprint arXiv:2407.16124 (2024)

    arXiv 2024

  23. [23]

    In: European Conference on Computer Vision (ECCV) (2024)

    Luvizon, D., Golyanik, V., Kortylewski, A., Habermann, M., Theobalt, C.: Re- lightable neural actor with intrinsic decomposition and pose control. In: European Conference on Computer Vision (ECCV) (2024)

    2024

  24. [24]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Mei, Y., He, M., Ma, L., Philip, J., Xian, W., George, D.M., Yu, X., Dedic, G., Taşel, A.L., Yu, N., et al.: Lux post facto: Learning portrait performance relight- ing with conditional video diffusion and a hybrid dataset. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5510–5522 (2025)

    2025

  25. [25]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Mei, Y., Zeng, Y., Zhang, H., Shu, Z., Zhang, X., Bi, S., Zhang, J., Jung, H., Patel, V.M.: Holo-relighting: Controllable volumetric portrait relighting from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4263–4273 (2024)

    2024

  26. [26]

    ACM Trans

    Pandey, R., Orts-Escolano, S., Legendre, C., Haene, C., Bouaziz, S., Rhemann, C., Debevec, P.E., Fanello, S.R.: Total relighting: learning to relight portraits for background replacement. ACM Trans. Graph.40(4), 43–1 (2021)

    2021

  27. [27]

    Acm transactions on graphics (tog)29(4), 1–13 (2010)

    Parker, S.G., Bigler, J., Dietrich, A., Friedrich, H., Hoberock, J., Luebke, D., McAl- lister, D., McGuire, M., Morley, K., Robison, A., et al.: Optix: a general purpose ray tracing engine. Acm transactions on graphics (tog)29(4), 1–13 (2010)

    2010

  28. [28]

    In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition (2024)

    Ren, M., Xiong, W., Yoon, J.S., Shu, Z., Zhang, J., Jung, H., Gerig, G., Zhang, H.: Relightful harmonization: Lighting-aware portrait background replacement. In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition (2024)

    2024

  29. [29]

    arXiv preprint arXiv:2201.02610 (2022)

    Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610 (2022)

    arXiv 2022

  30. [30]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Sanyal, S., Bolkart, T., Feng, H., Black, M.J.: Learning to regress 3d face shape and expression from an image without 3d supervision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7763–7772 (2019)

    2019

  31. [31]

    arXiv preprint arXiv:2512.00255 (2025)

    Singh, K.M., Chen, J., Golyanik, V., Garbin, S.J., Beeler, T., Dabral, R., Habermann, M., Theobalt, C.: Relightable holoported characters: Capturing and relighting dynamic human performance from sparse views. arXiv preprint arXiv:2512.00255 (2025)

    arXiv 2025

  32. [32]

    (2025),https://github.com/soravux/skylibs

    Skylibs: Tools used for ldr/hdr environment map (ibl) handling, conversion and i/o. (2025),https://github.com/soravux/skylibs

    2025

  33. [33]

    Team, T.H.: Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation (2025)

    2025

  34. [34]

    Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.W., Chen, D., Yu, F., Zhao, H., Yang, J., Zeng, J., Wang, J., Zhang, J., Zhou, J., Wang, J., Chen, J., Zhu, K., Generative Relightable Avatars 17 Zhao, K., Yan, K., Huang, L., Feng, M., Zhang, N., Li, P., Wu, P., Chu, R., Feng, R., Zhang, S., Sun, S., Fang, T., Wang, T., Gui, T., Weng, T., Shen, T., Lin,...

    Pith/arXiv arXiv 2025

  35. [35]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR) (2025)

    Wang, J., Liu, J., Sun, X., Singh, K.K., Shu, Z., Zhang, H., Yang, J., Zhao, N., Wang, T.Y., Chen, S.S., Neumann, U., Yoon, J.S.: Comprehensive relighting: Gen- eralizable and consistent monocular human relighting and harmonization. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR) (2025)

    2025

  36. [36]

    arXiv preprint arXiv:2509.23769 (2025)

    Wang, L., Jin, S., Cui, R., Dahl, A.B., Frisvad, J.R., Bigdeli, S.: Relumix: Ex- tending image relighting to video via video diffusion models. arXiv preprint arXiv:2509.23769 (2025)

    arXiv 2025

  37. [37]

    In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition

    Wang, S., Antic, B., Geiger, A., Tang, S.: Intrinsicavatar: Physically based inverse rendering of dynamic humans from monocular videos via explicit ray tracing. In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition. pp. 1877–1888 (2024)

    2024

  38. [38]

    In: Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers

    Wang, S., Simon, T., Santesteban, I., Bagautdinov, T., Li, J., Agrawal, V., Prada, F., Yu, S.I., Nalbone, P., Gramlich, M., et al.: Relightable full-body gaussian codec avatars. In: Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers. pp. 1–12 (2025)

    2025

  39. [39]

    In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision

    Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., Liu, L.: Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision. pp. 3295– 3306 (2023)

    2023

  40. [40]

    IEEE transactions on image processing13(4), 600–612 (2004)

    Wang, Z.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing13(4), 600–612 (2004)

    2004

  41. [41]

    Tech report (2025)

    Xiang, J., Chen, X., Xu, S., Wang, R., Lv, Z., Deng, Y., Zhu, H., Dong, Y., Zhao, H., Yuan, N.J., Yang, J.: Native and compact structured latents for 3d generation. Tech report (2025)

    2025

  42. [42]

    arXiv preprint arXiv:2412.01506 (2024)

    Xiang, J., Lv, Z., Xu, S., Deng, Y., Wang, R., Zhang, B., Chen, D., Tong, X., Yang, J.: Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506 (2024)

    Pith/arXiv arXiv 2024

  43. [43]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Xiao, J., Zhang, Q., Xu, Z., Zheng, W.S.: Neca: Neural customizable human avatar. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20091–20101 (2024)

    2024

  44. [44]

    ACM Transactions on Graphics43(5), 1–22 (2024)

    Xie, R., Huang, K., Cho, I.Y., Yang, S., Chen, W., Bao, H., Zheng, W., Li, R., Huo, Y.: Ren human: Learning relightable neural implicit surfaces for animatable human rendering. ACM Transactions on Graphics43(5), 1–22 (2024)

    2024

  45. [45]

    In: CVPR (2024)

    Xu, Z., Peng, S., Geng, C., Mou, L., Yan, Z., Sun, J., Bao, H., Zhou, X.: Relightable and animatable neural avatar from sparse-view video. In: CVPR (2024)

    2024

  46. [46]

    In: CVPR (2026)

    Yang, P., Zhou, S., Hao, K., Tao, Q.: MatAnyone 2: Scaling video matting via a learned quality evaluator. In: CVPR (2026)

    2026

  47. [47]

    ACM Transactions on Graphics (TOG) 43(6), 1–13 (2024)

    Yoon, J.S., Shu, Z., Ren, M., Zhang, C., Hold-Geoffroy, Y., Singh, K.k., Zhang, H.: Generative portrait shadow removal. ACM Transactions on Graphics (TOG) 43(6), 1–13 (2024)

    2024

  48. [48]

    Singh et al

    Zeng, Z., Deschaintre, V., Georgiev, I., Hold-Geoffroy, Y., Hu, Y., Luan, F., Yan, L.Q., Hašan, M.: Rgb-x: Image decomposition and synthesis using material- and 18 K.M. Singh et al. lighting-aware diffusion models. In: ACM SIGGRAPH 2024 Conference Papers (2024)

    2024

  49. [49]

    In: International Conference on Learning Representations (ICLR) (2025)

    Zhang, L., Rao, A., Agrawala, M.: Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In: International Conference on Learning Representations (ICLR) (2025)

    2025

  50. [50]

    In: CVPR (2018)

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)

    2018

  51. [51]

    Zhao, Y., Wu, C., Huang, B., Zhi, Y., Zhao, C., Wang, J., Gao, S.: Surfel-based gaussian inverse rendering for fast and relightable dynamic human reconstruction from monocular video. arXiv preprint arXiv:2407.15212 (2024) Generative Relightable Avatars 19 A Overview This supplementary document provides additional details and analyses to com- plement the m...

    arXiv 2024

  52. [52]

    C" denotes convolution layer,

    Table 3 illustrates the concrete architecture ofRelightNet. Table 3:Illustration of theRelightNetarchitecture. In the operation column, "C" denotes convolution layer, "SA" denotes self-attention layer, "CA" denotes cross- attention layer, "DS" and "US" denotes down-sampling and up-sampling layer with scale factors equal to 2. Number of Feature Channels Op...