Pith. sign in

REVIEW 4 major objections 5 minor 34 references

VideoMat: Extracting PBR Materials from Video Diffusion Models

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read VideoMat extracts high-quality, spatially varying PBR material maps (base color, roughness, metallic) for a given 3D model from a text prompt or single image, using a geometry- and lighting-conditioned video diffusion model, video…

desk verdict A genuinely new pipeline for PBR material extraction from video diffusion with a differentiable path tracer; solid evidence but the lighting-consistency assumption needs direct verification. read the letter →

arxiv 2506.09665 v2 pith:SSWQJSHL submitted 2025-06-11 cs.GR cs.CV

Jacob Munkberg , Zian Wang , Ruofan Liang , Tianchang Shen , Jon Hasselgren This is my paper
classification cs.GRcs.CV
keywords text-to-materialPBRmaterialsvideodiffusionmodelsdifferentiablepathtracingintrinsicdecompositionrelightingmaterialgeneration3Dtexturing
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 introduces VideoMat, a pipeline that turns a text prompt (or a single reference image) plus known 3D geometry into physically based rendering (PBR) material maps: base color, roughness, and metallic. The key move is to generate a short orbit video of the object with a finetuned video diffusion model conditioned on the geometry's normal maps and a controlled HDR light probe, rather than drawing independent views from image models. Because video generation keeps material properties coherent across frames, the subsequent intrinsic decomposition and multi-view optimization can separate lighting from surface appearance instead of baking highlights into the texture. The paper argues that this produces higher-quality, more relightable materials than existing text-to-material methods, which is what a practitioner would need for content creation.

What carries the argument

Two finetuned video diffusion models sandwich a physically based inverse renderer. Video Model I is a Cosmos DiT finetuned with extra input channels: camera-space normal maps plus a compressed RGB encoding of three uniform-material shading passes (roughness 1/0.5/0 and metallic 0/0.5/1) lit by the HDR probe, which tells the generator both geometry and lighting. Video Model II, re-implemented from DiffusionRenderer, outputs per-frame intrinsic G-buffers (base color/roughness/metallic) from the generated video. The differentiable path tracer, built like Mitsuba3 with hash grid+MLP material evaluation in the ray tracing loop, reconstructs spatially varying Disney-style PBR textures from the generated frames, regularized with a scale-invariant L1 loss on the predicted intrinsics.

What would settle it

Take a mesh with a known ground-truth PBR material, render orbit videos under a known HDR probe, and check whether the same extracted material re-lit with another probe reproduces the corresponding generated frames; if the extracted base color maps differ between two probes, lighting is still being baked in. A simpler direct test is to generate the same object with two different HDR probes and compare the extracted base color textures: material maps that resemble each other under different lighting indicate good separation, while systematic differences reveal residual entanglement.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a video diffusion model, finetuned to respect input normals and a lighting condition, can serve as a view-consistent generative prior for PBR material extraction. Given a mesh and an HDR probe, VideoMat renders normal and simple-shading guides, feeds them with a text prompt to the finetuned Cosmos model to synthesize a 121-frame orbit of the object with a novel material, then runs a second video diffusion model (re-implemented from DiffusionRenderer) to decompose each frame into base color, roughness, and metallic. A differentiable path tracer with an inline hash grid+MLP material representation then fits PBR textures to all 121 frames, with a scale-invariant regularizer that keeps the predicted intrinsics aligned. The paper claims this beats image-diffusion baselines such as DreamMat and Paint-it in separation of lighting and materials, with higher PSNR and LPIPS scores on relit renders and user-study preference.

Load-bearing premise

The whole pipeline depends on the finetuned video model actually producing frames that stay faithful to the supplied normal maps and lighting, with a material that stays coherent as the camera orbits; if the generated video drifts, the later decomposition and path-tracing optimization inherit that drift and cannot correctly separate lighting from material.

Editorial extensions

If this is right

  • Text-to-material generation from a single prompt and mesh yields spatially varying, relightable PBR maps that are directly usable in content creation tools.
  • Because the generator is conditioned on geometry and lighting, the extracted materials separate diffuse albedo from specular response, so relighting under novel HDR probes looks correct rather than reusing baked highlights.
  • The image-conditioned variant can perform material inpainting or style transfer from a single shaded view of a reference scene.
  • The method can serve as a material-generation component for text-to-3D and single-image reconstruction systems, and for augmenting 3D datasets with diverse materials.
  • Future improvements to video models, such as spherical camera coverage or reduced latent-space compression, should translate directly into higher-fidelity textures.

Reading between the lines

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

  • If video-model drift is the main failure mode, then any improvement in video diffusion consistency or camera control should directly raise material fidelity; this is a testable prediction the paper's ablation structure could check.
  • The scale-invariant regularizer compensates for bias in the intrinsic model, and the paper's ablations suggest quality drops without it, so a broader comparison of regularizer strength across diverse scenes would clarify how much of the gain comes from the video prior versus the optimization.
  • The same two-model-plus-path-tracer structure likely extends to other appearance channels such as normal mapping, refraction, and subsurface scattering, which the paper lists as future work.
  • Because the generated video is the only source of supervision, the method inherits the diversity and biases of the base video model; material types underrepresented in training data should produce less faithful results, a consequence worth probing with a broad material taxonomy.
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

4 major / 5 minor

Summary. The paper presents VideoMat, a pipeline for extracting spatially varying PBR material maps (base color, roughness, metallic) for a known 3D mesh from a text prompt or a single reference image. The method finetunes a Cosmos video diffusion model to generate a 121-frame 360-degree orbit video conditioned on normal maps, three simple shading guides, and text. A second video model, a re-implementation of the DiffusionRenderer inverse renderer, predicts per-frame intrinsic maps. A differentiable path tracer with a hash-grid and MLP material representation then optimizes the material maps against the generated video, regularized by the predicted intrinsics. Quantitative evaluation is performed on six held-out BlenderVault scenes against Paint-it, DreamMat, and Make-it-Real, with relighting and intrinsic metrics; qualitative results, ablations, and a small user study are also reported. The central claim is that the approach yields higher-quality PBR materials and better separation of lighting and material than prior image-diffusion-based methods.

Significance. If validated, the work is a meaningful step for text-to-material generation, since it combines view-consistent video diffusion with a physically based differentiable renderer and uses held-out BlenderVault ground truth to evaluate relative to prior methods. The ablations of the intrinsic regularizer and path-tracing depth are useful and the method description is mostly clear. The evaluation, however, is currently too thin to support the headline claims: six scenes with no error bars, a user study with eleven participants, and no direct validation that the generated video actually follows the input HDR lighting and geometry. Because the core disentanglement claim depends on this conditioning fidelity, the evidence presented does not yet fully establish the central result.

major comments (4)
  1. [Section 3.1, Section 5, Eq. (3)] The conditioning fidelity of the finetuned video model is never directly measured, and it is load-bearing for the central disentanglement claim. The normal and shading guides are encoded with the frozen Cosmos tokenizer with 8x8x8 compression in latent space (Section 3.1), and Section 5 concedes that aggressively compressed latent spaces produce motion blur and compression artifacts that 'break the pinhole camera and instant shutter assumptions' of the differentiable renderer. Since the L1 loss in Eq. (3) compares the renderer against the generated frames, any deviation from the known HDR probe and geometry (shifted highlights, softened shadows, invented fill light) cannot be represented by the rendering equation and will be absorbed into the material maps. The paper should provide a direct measurement of video fidelity, for example by generating videos from held-out BlenderVault scenes with known materials and probes and reporting frame-level PSNR or a shading-guide consistency metric against ground-truth renders.
  2. [Section 4, Table 1] The quantitative comparison in Table 1 lacks error bars, per-scene breakdowns, and significance tests. The table reports averages over six test scenes, and the text mentions one outlier (the Shed scene) without giving the corresponding per-scene numbers. Since the paper's claim of 'higher quality results and improved separation of lighting and materials' is supported primarily by this table, please provide per-scene metrics and confidence intervals, and run paired significance tests for the relit PSNR and base-color PSNR comparisons.
  3. [Section 4.1, Table 2] The user study is too small to support the perceptual superiority claim by itself. There are 11 participants, 24 samples, majority voting, and no statistical significance test; the GPT-4o result of 100% preference over Paint-it is especially fragile with this sample size. The authors should present per-sample results, use a binomial test or exact confidence intervals, and label the study as preliminary rather than conclusive evidence of perceptual quality.
  4. [Section 3.3, Eq. (4), Figure 5] The scale-invariant regularizer in Eq. (4) relies on Video Model II, which the paper itself describes as biased toward lower roughness (Figure 5 caption in Section 4). Because this regularizer is exactly what demodulates the albedo in the Figure 9 ablation, the paper should quantify Video Model II's intrinsic bias on the six BlenderVault scenes and demonstrate whether the final extracted materials inherit or correct that bias. Without this, the regularizer may be enforcing a learned bias rather than physical consistency.
minor comments (5)
  1. [Section 3.1] There is a typo: 'We use the the denoising score matching loss' should be 'We use the denoising score matching loss.'
  2. [Section 5] The sentence 'Optimization with an L1 loss recovers some image detail, but that a more principled solution is desired' is ungrammatical; consider rewriting to 'Optimization with an L1 loss recovers some image detail, but a more principled solution is desired.'
  3. [Section 4, Table 1] Please clarify whether the PSNR and LPIPS values are computed on tonemapped sRGB images, which pixels are included in the mask, and the exact number of relighting probes and frames used per scene.
  4. [Section 3.3, Eq. (4)] The notation x/x-bar in Eq. (4) is unclear; state explicitly that this is an element-wise division by the mean value of the image, and specify whether the mean is computed per channel.
  5. [Section 4] The text says the generated prompts are in the supplemental material; for reproducibility, include the full prompts for the six BlenderVault scenes in the main paper or in a clearly labeled appendix.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the extracted PBR maps are produced by a differentiable-rendering optimization against generated video and are independently evaluated on held-out BlenderVault ground truth.

full rationale

VideoMat's central extraction does not reduce to its inputs. The final material maps are obtained by solving the optimization in Eq. (3): argmin_phi E_c L(I_phi(c), I_ref(c)), where I_ref are frames generated by the finetuned Cosmos video model and L combines an L1 image term with a scale-invariant regularizer on Video Model II's intrinsics. The output maps are not defined in terms of the BlenderVault references; those references are used only for post hoc evaluation, with BlenderVault held out from Video Model I training, so the benchmark is independent. The conditioning on normals and shading guides is control input to the generative model, not the extracted material, and it follows prior work (DreamMat, FlashTex) rather than an unverified self-citation. The only self-referential element is Video Model II, a re-implementation of the authors' DiffusionRenderer (Sec. 3.2: 'we re-implement the inverse renderer from DiffusionRenderer'), which supplies the intrinsic regularizer. That regularizer is a learned prior trained on 150k Objaverse videos, is ablated in Fig. 9, and is assigned weight lambda = 0.2; it influences the result but does not constitute it, and no uniqueness theorem or untested ansatz is imported from the citation. The limitations in Sec. 5—latent-space compression breaking the pinhole-camera and instant-shutter assumptions—are genuine correctness risks, since drift in the generated video could let the L1 optimization bake lighting into materials, but that is an empirical robustness concern rather than a circular derivation. Overall, the pipeline is self-contained and externally validated, so no circularity is present.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The central claim rests on two learned models and a set of engineering hyperparameters rather than on a new conservation law or mathematical object. The main scientific assumptions are that the finetuned video model respects geometric and lighting conditions and that the intrinsic-decomposition model supplies trustworthy guides.

free parameters (3)
  • regularization weight lambda = 0.2
    Chosen empirically in Section 3.3 to balance the image reconstruction loss and the three intrinsic regularizer terms; the stated value works well over the test scenes but is not derived from first principles.
  • shading guide material parameters = kd=0.7, r=1/0.5/0, m=0/0.5/1
    Hand-selected conditions in Section 3.1 used to encode lighting into the video model conditioning; these choices affect how the model interprets the lighting condition.
  • optimization hyperparameters = 1k iterations, batch size 8, 128 spp forward, 4 spp backward, 1-3 bounces
    Computational settings in Section 3.3 chosen to make reconstruction robust and tractable; they are not fitted to the evaluation data but are free choices affecting quality.
assumptions (5)
  • domain assumption The input includes known 3D geometry with a unique texture parametrization and an HDR light probe.
    Stated at the start of Section 3; the entire pipeline depends on geometry and lighting being available.
  • domain assumption The Cosmos base model can be finetuned to respect normal-map and shading conditions while generating coherent novel materials.
    Section 3.1 assumes 10k iterations on a 40k-video dataset is sufficient; the paper provides qualitative evidence but no formal guarantee.
  • domain assumption The DiffusionRenderer-style inverse renderer produces reliable video intrinsics for base color, roughness, and metallic.
    Section 3.2 adopts and finetunes this inverse renderer; the regularization in Section 3.3 treats its outputs as trustworthy guides.
  • domain assumption The Disney PBR BSDF and a path tracer with 1 to 3 bounces sufficiently model the generated videos.
    Section 3.3 uses this shading model; effects like refraction, subsurface scattering, and normal maps are explicitly left out in Section 5.
  • domain assumption Six BlenderVault scenes held out from training are representative enough to compare text-to-material methods.
    Section 4 uses these six scenes for all quantitative metrics; no cross-dataset evaluation or variance analysis is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VideoMat: Extracting PBR Materials from Video Diffusion Models." pith.science (2026). https://pith.science/paper/SSWQJSHL

@misc{pith2026250609665,
  author       = {Pith},
  title        = {Pith review of: VideoMat: Extracting PBR Materials from Video Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SSWQJSHL}},
  note         = {Machine review of arXiv:2506.09665}
}
read the original abstract

We leverage finetuned video diffusion models, intrinsic decomposition of videos, and physically-based differentiable rendering to generate high quality materials for 3D models given a text prompt or a single image. We condition a video diffusion model to respect the input geometry and lighting condition. This model produces multiple views of a given 3D model with coherent material properties. Secondly, we use a recent model to extract intrinsics (base color, roughness, metallic) from the generated video. Finally, we use the intrinsics alongside the generated video in a differentiable path tracer to robustly extract PBR materials directly compatible with common content creation tools.

Figures

Figures reproduced from arXiv: 2506.09665 by the authors.

Figure 1
Figure 1. Given a 3D model and a prompt, we extract high quality PBR materials from finetuned video diffusion models. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Our method starts from a known 3D model and an HDR environment map. We first render videos of normal maps and three simple uniform shading [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. We condition the video model with normal maps and three shading [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Our image-to-material pipeline is conditioned on a single image, [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Text-to-material generation. We compare against Paint-it [Youwang et al [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 7
Figure 7. Figure 7: Our finetuned video model generates view-consistent multi-view [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 6
Figure 6. Figure 6: The Helmet model rendered with four light probes chosen to stress￾test different lighting conditions. Our result show accurate specular reflec￾tions. DreamMat reproduces the global color tone of the probe, but the baked highlights are static and do not reflect the prim…
Figure 10
Figure 10. Figure 10: Impact of physically-based rendering. In this example, a significant [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]
Figure 9
Figure 9. Figure 9: The impact of using intrinsic guides for regularization. Differentiable [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 18 canonical work pages

  1. [3]

    arXiv preprint arXiv:2409.12957 (2024)

    3DTopia-XL: High-Quality 3D PBR Asset Generation via Primitive Diffusion. arXiv preprint arXiv:2409.12957 (2024). Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik

    arXiv 2024

  2. [6]

    arXiv:2404.16829 [cs.CV] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin- Brualla, Pratul P

    Make-it-Real: Unleashing Large Multimodal Model for Painting 3D Objects with Realistic Materials. arXiv:2404.16829 [cs.CV] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin- Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*

    arXiv

  3. [7]

    Advances in Neural Information Processing Systems (2024)

    CAT3D: Create Anything in 3D with Multi-View Diffusion Models. Advances in Neural Information Processing Systems (2024). Eric Heitz

    work page 2024

  4. [10]

    arXiv preprint arXiv:2408.16500 (2024)

    CogVLM2: Visual Language Models for Image and Video Understanding. arXiv preprint arXiv:2408.16500 (2024). Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang

    arXiv 2024

  5. [12]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Transactions on Graphics 42, 4 (July 2023). https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/ Mukul Khanna*, Yongsen Mao*, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X. Chang, and Manolis Savva

    work page 2023

  6. [13]

    arXiv preprint (2023)

    Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation. arXiv preprint (2023). arXiv:2306.11290 [cs.CV] Diederik P. Kingma and Jimmy Ba

    arXiv 2023

  7. [16]

    ACM Transactions on Graphics 39, 6 (2020)

    Modular Primitives for High-Performance Differentiable Rendering. ACM Transactions on Graphics 39, 6 (2020). Ruofan Liang, Zan Gojcic, Huan Ling, Jacob Munkberg, Jon Hasselgren, Zhi-Hao Lin, Jun Gao, Alexander Keller, Nandita Vijaykumar, Sanja Fidler, and Zian Wang

    work page 2020

  8. [17]

    In SIGGRAPH Asia 2023 Technical Communications (SA ’23)

    Standard Shader Ball: A Modern and Feature- Rich Render Test Scene. In SIGGRAPH Asia 2023 Technical Communications (SA ’23) . Article 14, 3 pages. Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ra- mamoorthi, and Ren Ng

    work page 2023

Show all 34 references
  1. [18]

    ACM Trans

    Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Trans. Graph. 41, 4 (2022), 102:1–102:15. Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler

    2022

  2. [19]

    arXiv preprint arXiv:2501.03575 (2025)

    Cosmos World Foundation Model Platform for Physical AI. arXiv preprint arXiv:2501.03575 (2025). OpenAI

    2025 arXiv

  3. [20]

    arXiv:2410.21276 https://arxiv.org/abs/2410.21276 Steven G

    GPT-4o System Card. arXiv:2410.21276 https://arxiv.org/abs/2410.21276 Steven G. Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David Luebke, David McAllister, Morgan McGuire, Keith Morley, Austin Robison, and Martin Stich

    arXiv

  4. [22]

    arXiv (2022)

    DreamFusion: Text-to-3D using 2D Diffusion. arXiv (2022). Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier- David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao

    2022

  5. [23]

    In ACM SIGGRAPH 2023 conference proceedings

    TEXTure: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings. 1–11. Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman

    2023

  6. [24]

    arXiv preprint arxiv:2208.12242 (2022)

    DreamBooth: Fine Tuning Text-to-image Diffusion Models for Subject-Driven Generation. arXiv preprint arxiv:2208.12242 (2022). Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang

    2022 arXiv

  7. [25]

    arXiv:2308.16512 (2023)

    MV- Dream: Multi-view Diffusion for 3D Generation. arXiv:2308.16512 (2023). Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli

    2023 arXiv

  8. [26]

    In Advances in Neural Information Processing Systems (NeurIPS)

    ProlificDreamer: High-Fidelity and Diverse Text-to-3D Generation with Variational Score Distillation. In Advances in Neural Information Processing Systems (NeurIPS). Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T. Barron, and Aleksander Holynski. 20...

    2024 arXiv

  9. [27]

    arxiv: 2404.11593 (2024)

    IntrinsicAnything: Learning Diffusion Priors for Inverse Rendering Under Unknown Illumination. arxiv: 2404.11593 (2024). Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang

    2024 arXiv

  10. [28]

    arXiv preprint arXiv:2412.01506 (2024)

    Structured 3D Latents for Scalable and Versatile 3D Generation. arXiv preprint arXiv:2412.01506 (2024). Haibo Yang, Yang Chen, Yingwei Pan, Ting Yao, Zhineng Chen, Chong-Wah Ngo, and Tao Mei. 2024a. Hi3D: Pursuing High-Resolution Image-to-3D Generation with Video Diffusion Mod...

    2024 arXiv

  11. [29]

    arXiv preprint arXiv:2401.09416 (2024)

    Tex- tureDreamer: Image-guided Texture Synthesis through Geometry-aware Diffusion. arXiv preprint arXiv:2401.09416 (2024). Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll

    2024 arXiv

  12. [30]

    ACM Trans

    TEXGen: a Generative Diffusion Model for Mesh Textures. ACM Trans. Graph. 43, 6, Article 213 (2024). Greg Zaal and et al

    2024

  13. [31]

    https://polyhaven

    Poly Haven - The Public 3D Asset Library . https://polyhaven. com Chong Zeng, Yue Dong, Pieter Peers, Youkang Kong, Hongzhi Wu, and Xin Tong. 2024c. DiLightNet: Fine-grained Lighting Control for Diffusion-based Image Generation. In ACM SIGGRAPH 2024 Conference Papers . Xianfan...

    2024

  14. [32]

    ACM Trans

    Path-Space Differentiable Rendering. ACM Trans. Graph. 39, 4 (2020), 143:1–143:19. Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. 2024c. CLAY: A Controllable Large-scale Generative Model for Creating High-quality 3D...

    2020

  15. [33]

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

    The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition . 586–595. Shangzhan Zhang, Sida Peng, Tao Xu, Yuanbo Yang, Tianrun Chen, Nan Xue, Yujun Shen, Hujun Bao, Ruizhen Hu, and Xi...

    2024

  16. [34]

    HiFA: High-fidelity Text-to-3D Generation with Advanced Diffusion Guidance. arXiv:2305.18766 [cs.CV] Shenhao Zhu, Lingteng Qiu, Xiaodong Gu, Zhengyi Zhao, Chao Xu, Yuxiao He, Zhe Li, Xiaoguang Han, Yao Yao, Xun Cao, Siyu Zhu, Weihao Yuan, Zilong Dong, and Hao Zhu

    arXiv

  17. [1982]

    ACM Transactions on Graphics (ToG) 1, 1 (1982), 7–24

    A reflectance model for computer graphics. ACM Transactions on Graphics (ToG) 1, 1 (1982), 7–24. Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli Vander- Bilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi

    1982

  18. [2010]

    ACM Trans

    OptiX: A general purpose ray tracing engine. ACM Trans. Graph. 29, 4, Article 66 (2010), 13 pages. Matt Pharr, Wenzel Jacob, and Greg Humphreys

    2010

  19. [2013]

    Real shading in Unreal Engine 4.ACM SIGGRAPH Course on Physically Based Shading Theory and Practice 4, 3 (2013),

    2013

  20. [2017]

    arXiv:1412.6980 [cs.LG] Peter Kocsis, Lukas Höllein, and Matthias Nießner

    Adam: A Method for Stochastic Optimization. arXiv:1412.6980 [cs.LG] Peter Kocsis, Lukas Höllein, and Matthias Nießner

    arXiv

  21. [2018]

    Journal of Computer Graphics Techniques (JCGT) 7, 4 (2018), 1–13

    Sampling the GGX Distribution of Visible Normals. Journal of Computer Graphics Techniques (JCGT) 7, 4 (2018), 1–13. Jonathan Ho, Ajay Jain, and Pieter Abbeel

    2018

  22. [2020]

    Advances in Neural Information Processing Systems 33 (2020), 6840–6851

    Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33 (2020), 6840–6851. Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang

    2020

  23. [2022]

    CVPR (2022)

    ABO: Dataset and Benchmarks for Real-World 3D Object Understanding. CVPR (2022). Robert L. Cook and Kenneth E. Torrance

    2022

  24. [2023]

    ACM Trans

    SLANG.D: Fast, Modular and Differentiable Shader Programming. ACM Trans. Graph. 42, 6, Article 264 (2023). Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and R...

    2023 arXiv

  25. [2024]

    arXiv:2404.02899 https://arxiv.org/abs/2404.02899 Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner

    MatAtlas: Text- driven Consistent Geometry Texturing and Material Assignment. arXiv:2404.02899 https://arxiv.org/abs/2404.02899 Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. 2023b. Text2tex: Text-driven texture synthesis via diffusion ...

    arXiv

  26. [2025]

    arXiv:2504.01008 [cs.CV] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila

    IntrinsiX: High-Quality PBR Generation using Image Priors. arXiv:2504.01008 [cs.CV] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.