Pith. sign in

REVIEW 3 major objections 5 minor 59 references

Φeat learns material identity from renderings alone, no labels needed.

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

2026-08-03 22:13 UTC pith:H66HYCCN

load-bearing objection Φeat's core idea—physical render variation as contrastive augmentation—is genuinely new, but the headline k-NN gains are undermined by a missing held-out split and an inconsistent DINOv3 baseline. the 3 major comments →

arxiv 2511.11270 v2 pith:H66HYCCN submitted 2025-11-14 cs.CV

{Φ}eat: Physically Grounded Material Feature Representation

Giuseppe Vecchio , Adrien Kaiser , Claudia Cuttano , Rouffet Romain , Rosalie Martin , Elena Garces , Tamy Boubekeur This is my paper
classification cs.CV
keywords self-supervised learningmaterial representationreflectancephysical invariancevision transformercontrastive learningmaterial selectionsynthetic rendering
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.

Φeat is a self-supervised vision backbone trained to make material identity the organizing principle of its features. The key move is replacing photometric augmentations with physically based renderings: the model sees the same material on different geometries and under different lighting and is trained to pull those views together, while an in-batch contrastive term pushes different materials apart. Starting from a semantics-heavy ViT, this fine-tuning preserves global context while making patch and image features sensitive to reflectance, mesostructure, and transparency rather than object category. On material selection and k-NN material classification, Φeat outperforms strong semantic baselines, suggesting that physical invariance can be learned without any labels.

Core claim

The central claim is that a pure self-supervised objective, built from renderings of the same material under varying geometry and illumination, yields a visual representation that is invariant to extrinsic physical factors while preserving sensitivity to intrinsic material properties. The paper demonstrates this by fine-tuning a DINOv3-initialized ViT with a multi-render scheme and a contrastive loss, then evaluating frozen features. Φeat raises IoU on DuMaS material selection from 0.599 (DINOv3) to 0.776, raises k-NN top-1 from 60.0% to 64.3%, and lowers prediction variability under illumination and geometry changes. The ablations show that single-render fine-tuning improves material groupi

What carries the argument

The load-bearing mechanism is 'physical augmentation': each training pair consists of two physically based renderings of the same material on different geometry templates under different environment lighting, so the only stable identity across views is the material itself. The training keeps the DINOv3 teacher–student self-distillation (Sinkhorn–Knopp prototype assignments, masked patch reconstruction, KoLeo dispersion, Gram anchoring) and adds an in-batch InfoNCE contrastive term on global student embeddings that pulls together views of the same material and pushes apart different materials. The contrastive term is what recovers global separability after material-supervised fine-tuning.

Load-bearing premise

The central claim stands on the assumption that the synthetic test set is an honest out-of-sample measure — the paper does not state that the 972 test materials were held out from the ~36,000 pretraining material sets.

What would settle it

Compare the material IDs or procedural presets of the 972 test materials against the ~36,000 training material sets; any overlap would undermine the generalization claim. Alternatively, rerun the k-NN and robustness evaluations on a strictly disjoint set of new materials and on real photographs with known material labels; if accuracy falls to baseline levels, the learned invariance is not real.

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

If this is right

  • Frozen Φeat features group image patches by shared reflectance and texture rather than object semantics, improving fine-grained material selection.
  • Material k-NN classification improves with a balanced precision–recall profile, reaching 64.3% top-1 on a controlled synthetic test set.
  • Predictions are more stable across illumination and geometry changes, so fewer label flips occur when the same material is re-rendered.
  • The pretraining recipe does not need semantic labels and can reuse existing semantic backbones as initialization.
  • Ablations indicate that multi-render visibility plus the contrastive term is the combination that keeps features both physically grounded and globally discriminative.

Where Pith is reading between the lines

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

  • The paper does not state whether the 972 test materials were held out from the roughly 36,000 material sets used in pretraining; if they overlap, the k-NN and robustness numbers could partly reflect retrieval of familiar assets. A strictly disjoint test set would settle this.
  • The same physical-augmentation idea could be extended to disentangle other intrinsic factors, for example separating albedo from shading, or to learn invariance to camera viewpoint for material capture pipelines.
  • Because the model is pretrained only on synthetic renders, applying it to unpaired real photographs remains an open step; the paper itself acknowledges this in its limitations section.
  • The contrastive term's role suggests that pure grouping of physically equivalent views is not enough — without a push-apart force, features collapse in global discriminability; this interplay may guide future material-aware SSL designs.

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

3 major / 5 minor

Summary. The paper proposes Φeat, a vision transformer backbone initialized from DINOv3 and fine-tuned with a self-supervised objective that replaces photometric augmentations with physically meaningful ones: renderings of the same material on different geometries and under different lighting conditions are contrasted. The training combines DINOv3-style image-level, patch-level, KoLeo, and Gram-anchoring losses with an in-batch InfoNCE contrastive term. The method is evaluated on patch-level material selection on the DuMaS dataset and on a synthetic k-NN classification benchmark with robustness measurements, and is compared against CLIP, DINOv2, and DINOv3. The central claim is that this physically grounded pretraining yields representations that are invariant to lighting and geometry while preserving material identity, without using explicit labels.

Significance. If the claims are substantiated, the paper makes a useful contribution: the idea of replacing photometric augmentations with controlled renderings of the same material is a simple and scalable way to inject physical invariance into a self-supervised backbone, and the reported gains on material selection (F1 0.724 to 0.860 on DuMaS) are substantial. The method does not directly optimize the evaluation metrics, so the results are not circular by construction. However, the current evidence is not fully convincing: the synthetic k-NN test set may overlap with pretraining assets, the baseline numbers are inconsistent across tables, and no variance estimates are provided. The paper does not release code, so reproducibility rests entirely on the written description.

major comments (3)
  1. [Section 5.2, Tables 2–3] The k-NN test set is described as 'a controlled synthetic test set containing 972 materials within 16 categories,' but the paper never states that these materials, geometry templates, or environment maps were held out from the ~36,000 pretraining material sets described in Section 3. Since both training and test data are drawn from the Adobe Substance 3D Assets library, the Top-1 and robustness numbers may measure retrieval of familiar asset-specific BRDF/texture signatures rather than generalization to unseen materials. Please state explicitly whether the test materials are disjoint from the pretraining set; if they are not, re-run on a held-out library and report the resulting numbers.
  2. [Table 4 vs. Table 2] For the same DINOv3 baseline, the k-NN Top-1 accuracy is 0.600 in Table 2 but 0.563 in Table 4. The paper gives no explanation for this discrepancy. If the evaluation protocols differ (e.g., different splits, resolutions, or k values), the tables are not directly comparable and the comparisons are not controlled. If the protocols are identical, one of the numbers is wrong. Please align all baseline rows to a single evaluation protocol and report the corrected numbers in both tables.
  3. [Section 5.2] No error bars, seeds, or repeated runs are reported. The headline k-NN improvement (0.643 vs. 0.600 for DINOv3) is small, and with a single run it is impossible to judge whether it is statistically significant, especially given the baseline inconsistency noted above. Please report means and standard deviations across at least three seeds or bootstrap resamples for both k-NN and robustness metrics.
minor comments (5)
  1. [Section 3] The text says the pipeline yields 'approximately 36,000 unique sets of PBR maps' and later 'roughly one million high-quality renders.' Please clarify the relationship between these numbers (e.g., number of material sets × geometries × lighting conditions).
  2. [Section 5.2] Typo: 'exibiting' should be 'exhibiting' in the robustness paragraph.
  3. [Table 3 caption] The caption refers to 'the same underlying asset,' but the text and evaluation describe the same material under different lighting and geometry. Please use consistent terminology.
  4. [Section 5.4] The ablation would be clearer if the table also reported the robustness Hamming distances for the ablated variants, since Tables 2 and 3 together support the claim that the contrastive term improves separability and invariance.
  5. [Section 6] The limitations paragraph correctly notes that the method does not disentangle latent factors and relies on synthetic data. It would be helpful to also mention the absence of real-image evaluation, which is currently only qualitative.

Circularity Check

0 steps flagged

No circular derivation; central training/evaluation are aligned but not self-justifying, with minor self-citation and an unstated test-set split.

full rationale

Φeat's derivation is a standard self-supervised fine-tuning pipeline: renderings of the same material under varied geometry/lighting define positives, DINO/iBOT losses plus an in-batch InfoNCE term shape the embedding. The evaluation metrics (material selection on DuMaS, k-NN on synthetic renders, Hamming robustness) are computed on frozen encoders with fixed thresholds and k; none of the reported numbers is a fitted parameter renamed as a prediction, and no equation in Sec. 4 reduces to the metrics in Tables 1-3. The only mild self-citation is the use of DuMaS [27] for material selection; one co-author of the present paper (Elena Garces) is also a DuMaS author, but DuMaS is an externally defined benchmark with its own ground-truth labels, so this does not make the comparison circular. A more substantive validity risk, though not a circularity, is that Sec. 5.2 never states that the 972 k-NN test materials are disjoint from the ~36,000 pretraining material sets from the same Adobe Substance 3D Assets library; if they overlap, the k-NN and robustness numbers would measure memorization rather than generalization. The discrepancy between DINOv3 Top-1 in Table 2 (0.600) and Table 4 (0.563) also suggests the evaluation protocol is not fully pinned down. These are correctness/leakage concerns, not evidence that the derivation is equivalent to its inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on training data and hyperparameters rather than a derivation. The hand-set hyperparameters are modest (loss weights, temperature, mask schedule, threshold), and the largest implicit assumptions are that synthetic renderings transfer to real data and that semantically constrained shape–material pairing is not a confound. No new physical entities, forces, or dimensions are introduced.

free parameters (4)
  • Loss weights λp, λk, λg, λc = 1.0, 0.1, 0.7, 0.25
    Hand-selected in Section 5.1; the balance of image, patch, KoLeo, Gram, and contrastive losses directly shapes the learned features.
  • InfoNCE temperature τ = 0.1
    Set in Section 5.1; controls the sharpness of the material contrastive grouping.
  • Teacher momentum schedule = Cosine 0.996 → 1.0
    DINOv3-style schedule in Section 5.1; affects training stability and feature quality.
  • Material-selection binarization threshold = 0.5
    Used in Section 5.2 to convert cosine-similarity maps into masks; changing this threshold changes IoU/F1 and is not swept across models.
axioms (4)
  • domain assumption Renderings of the same procedural material under changed geometry and lighting define the correct equivalence class for material identity.
    Section 3/4; if the material is not visually recognizable across these variations, the contrastive signal is mislabeled and the learned invariance is misguided.
  • domain assumption Semantically aligned pairing of material and macro-geometry does not introduce a shape/material confound.
    Section 3 states cork is rendered on rigid, low-curvature surfaces rather than wrinkled cloth; this may let the model use geometry as a proxy for material class.
  • domain assumption Synthetic PBR renderings transfer sufficiently well to real images.
    All pretraining is synthetic; Section 6 concedes 'closing the domain gap to unpaired real photographs remains an open challenge.'
  • standard math DINOv3 teacher–student distillation, Sinkhorn–Knopp, KoLeo, and iBOT objectives behave as described in the cited prior work.
    Section 4 relies on formulations from [11,48,58] without re-deriving them.

reviewed 2026-08-03 · how reviews work

0 comments
read the original abstract

While foundation models have emerged as general-purpose visual backbones, their representations are primarily optimized for semantics and lack explicit modeling of physical factors, such as reflectance, hindering their efficacy in tasks requiring explicit material reasoning. We introduce $\Phi$eat$, a novel material-grounded visual backbone that encourages a representation sensitive to material identity, including reflectance and mesostructure. Instead of relying on generic data augmentations, we pretrain our model by contrasting observations of the same material under controlled variations in lighting and geometry. This encourages invariance to extrinsic factors while preserving sensitivity to intrinsic material properties. We show that the resulting representation provides strong priors for material-centric tasks, including feature-based material selection and classification. Our results demonstrate that physically inspired weak supervision is an effective strategy for learning representations tailored to material perception.

Figures

Figures reproduced from arXiv: 2511.11270 by Adrien Kaiser, Claudia Cuttano, Elena Garces, Giuseppe Vecchio, Rosalie Martin, Rouffet Romain, Tamy Boubekeur.

Figure 1
Figure 1. Figure 1: We present Φeat, a novel physically-grounded foundation model sensitive to the physical properties that govern real-world appearance such as geometry, reflectance, transparency, and lighting. We visualize the cosine similarity maps obtained with Φeat output features between the patches marked with a red cross and all other patches. Abstract Foundation models have emerged as effective backbones for many vis… view at source ↗
Figure 2
Figure 2. Figure 2: Dataset samples. Each row shows a different template, first with no materials applied (first column) and then different materials (following columns). Materials are rendered with trans￾parency on objects matching their semantic categories, and ran￾domized object and lighting rotation. 3. Data Collection and Curation The effectiveness of self-supervised pretraining depends strongly on how the input variabil… view at source ↗
Figure 3
Figure 3. Figure 3: Φeat training pipeline. Two renderings of the same material are sampled and augmented with a multi-crop strategy that yields global and local views. The student processes all crops (globals and locals), with random masking on patch tokens for latent reconstruction; the teacher and the Gram teacher process global crops only. Both networks output class- and patch-level embeddings. At the image level, Sinkhor… view at source ↗
Figure 4
Figure 4. Figure 4: Patch-wise similarity and unsupervised segmentation. Left group shows cosine similarity maps between the embedding of a reference patch (red cross) and all others, visualizing the spatial coherence of learned representations. We show examples gradually growing from a mostly flat surface to a medium scale scene. Right group displays K-means segmentations obtained from the patch embeddings. Compared to DINOv… 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

59 extracted references · 9 linked inside Pith

  1. [1]

    Substance 3D Assets.https://substance3d

    Adobe. Substance 3D Assets.https://substance3d. adobe.com/assets, 2025. 4

    2025

  2. [2]

    Self-supervised learning from images with a joint-embedding predictive architecture

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bo- janowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619–15629, 2023. 3

    2023

  3. [3]

    Efficient self-supervised learning with contextualized target representations for vision, speech and language

    Alexei Baevski, Arun Babu, Wei-Ning Hsu, and Michael Auli. Efficient self-supervised learning with contextualized target representations for vision, speech and language. InIn- ternational Conference on Machine Learning, pages 1416–

  4. [4]

    Beit: Bert pre-training of image transformers.arXiv preprint arXiv:2106.08254, 2021

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers.arXiv preprint arXiv:2106.08254, 2021. 3

    Pith/arXiv arXiv 2021

  5. [5]

    Re- covering intrinsic scene characteristics.Comput

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

    1978

  6. [6]

    Intrinsic images in the wild.ACM Transactions on Graphics (TOG), 33(4): 1–12, 2014

    Sean Bell, Kavita Bala, and Noah Snavely. Intrinsic images in the wild.ACM Transactions on Graphics (TOG), 33(4): 1–12, 2014. 2

    2014

  7. [7]

    An l 1 image trans- form for edge-preserving smoothing and scene-level intrinsic decomposition.ACM Transactions On Graphics (TOG), 34 (4):1–12, 2015

    Sai Bi, Xiaoguang Han, and Yizhou Yu. An l 1 image trans- form for edge-preserving smoothing and scene-level intrinsic decomposition.ACM Transactions On Graphics (TOG), 34 (4):1–12, 2015. 2

    2015

  8. [8]

    Matclip: Light-and shape-insensitive assignment of pbr material models

    Michael Birsak, John Femiani, Biao Zhang, and Peter Wonka. Matclip: Light-and shape-insensitive assignment of pbr material models. InProceedings of the Special Inter- est Group on Computer Graphics and Interactive Techniques Conference Conference Papers, pages 1–10, 2025. 3

    2025

  9. [9]

    Extending the disney brdf to a bsdf with in- tegrated subsurface scattering

    Brent Burley. Extending the disney brdf to a bsdf with in- tegrated subsurface scattering. SIGGRAPH 2015 Course: Physically Based Shading in Theory and Practice, 2015. 4

    2015

  10. [10]

    Intrinsic image decomposi- tion via ordinal shading.ACM Transactions on Graphics, 43 (1):1–24, 2023

    Chris Careaga and Ya ˘gız Aksoy. Intrinsic image decomposi- tion via ordinal shading.ACM Transactions on Graphics, 43 (1):1–24, 2023. 2

    2023

  11. [11]

    Unsupervised learning of visual features by contrasting cluster assignments.Ad- vances in neural information processing systems, 33:9912– 9924, 2020

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Pi- otr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments.Ad- vances in neural information processing systems, 33:9912– 9924, 2020. 3, 5

    2020

  12. [12]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. InPro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021. 2, 3

    2021

  13. [13]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge- offrey Hinton. A simple framework for contrastive learning of visual representations. InInternational conference on ma- chine learning, pages 1597–1607. PmLR, 2020. 3

    2020

  14. [14]

    Exploring simple siamese rep- resentation learning

    Xinlei Chen and Kaiming He. Exploring simple siamese rep- resentation learning. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 15750–15758, 2021. 3

    2021

  15. [15]

    Vision transformers need registers.arXiv preprint arXiv:2309.16588, 2023

    Timoth ´ee Darcet, Maxime Oquab, Julien Mairal, and Pi- otr Bojanowski. Vision transformers need registers.arXiv preprint arXiv:2309.16588, 2023. 2, 3, 6

    Pith/arXiv arXiv 2023

  16. [16]

    Single-image svbrdf cap- ture with a rendering-aware deep network.ACM Transac- tions on Graphics (ToG), 37(4):1–15, 2018

    Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf cap- ture with a rendering-aware deep network.ACM Transac- tions on Graphics (ToG), 37(4):1–15, 2018. 2

    2018

  17. [17]

    Flexible SVBRDF capture with a multi-image deep network

    Valentin Deschaintre, Miika Aittala, Fr ´edo Durand, George Drettakis, and Adrien Bousseau. Flexible SVBRDF capture with a multi-image deep network. InComputer Graphics Forum, pages 1–13. Wiley Online Library, 2019. 2

    2019

  18. [18]

    The visual language of fabrics.ACM Transactions on Graphics (TOG), 42(4):1–15,

    Valentin Deschaintre, Julia Guerrero-Viu, Diego Gutierrez, Tamy Boubekeur, and Belen Masia. The visual language of fabrics.ACM Transactions on Graphics (TOG), 42(4):1–15,

  19. [19]

    Unsuper- vised visual representation learning by context prediction

    Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsuper- vised visual representation learning by context prediction. In Proceedings of the IEEE international conference on com- puter vision, pages 1422–1430, 2015. 3

    2015

  20. [20]

    Are large-scale datasets necessary for self-supervised pre-training?arXiv preprint arXiv:2112.10740, 2021

    Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Herv´e Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training?arXiv preprint arXiv:2112.10740, 2021. 3

    Pith/arXiv arXiv 2021

  21. [21]

    Intrinsic images by entropy minimization

    Graham D Finlayson, Mark S Drew, and Cheng Lu. Intrinsic images by entropy minimization. InEuropean conference on computer vision, pages 582–595. Springer, 2004. 2

    2004

  22. [22]

    Deep inverse rendering for high-resolution svbrdf estimation from an arbitrary number of images.ACM Trans

    Duan Gao, Xiao Li, Yue Dong, Pieter Peers, Kun Xu, and Xin Tong. Deep inverse rendering for high-resolution svbrdf estimation from an arbitrary number of images.ACM Trans. Graph., 38(4):134–1, 2019. 2

    2019

  23. [23]

    Intrinsic images by clustering

    Elena Garces, Adolfo Munoz, Jorge Lopez-Moreno, and Diego Gutierrez. Intrinsic images by clustering. InCom- puter graphics forum, pages 1415–1424. Wiley Online Li- brary, 2012. 2

    2012

  24. [24]

    A Survey on Intrinsic Images: Delv- ing Deep into Lambert and Beyond.International Journal of Computer Vision, 130(3):836–868, 2022

    Elena Garces, Carlos Rodriguez-Pardo, Dan Casas, and Jorge Lopez-Moreno. A Survey on Intrinsic Images: Delv- ing Deep into Lambert and Beyond.International Journal of Computer Vision, 130(3):836–868, 2022. 2

    2022

  25. [25]

    Towards material digitization with a dual-scale op- tical system.ACM Transactions on Graphics (TOG), 42(4): 1–13, 2023

    Elena Garces, Victor Arellano, Carlos Rodriguez-Pardo, David Pascual-Hernandez, Sergio Suja, and Jorge Lopez- Moreno. Towards material digitization with a dual-scale op- tical system.ACM Transactions on Graphics (TOG), 42(4): 1–13, 2023. 3

    2023

  26. [26]

    Bootstrap your own latent-a new approach to self-supervised learning.Advances in neural information processing systems, 33:21271–21284, 2020

    Jean-Bastien Grill, Florian Strub, Florent Altch ´e, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Ghesh- laghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning.Advances in neural information processing systems, 33:21271–21284, 2020. 3

    2020

  27. [27]

    Fine-grained spatially varying material selection in images.arXiv preprint arXiv:2506.09023, 2025

    Julia Guerrero-Viu, Michael Fischer, Iliyan Georgiev, Elena Garces, Diego Gutierrez, Belen Masia, and Valentin De- schaintre. Fine-grained spatially varying material selection in images.arXiv preprint arXiv:2506.09023, 2025. 2, 3, 6

    Pith/arXiv arXiv 2025

  28. [28]

    Momentum contrast for unsupervised visual rep- 9 resentation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual rep- 9 resentation learning. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 9729–9738, 2020. 3

    2020

  29. [29]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000– 16009, 2022. 2, 3

    2022

  30. [30]

    Determining lightness from an image

    Berthold KP Horn. Determining lightness from an image. Computer graphics and image processing, 3(4):277–299,

  31. [31]

    In- trinsic image diffusion for indoor single-view material es- timation

    Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. In- trinsic image diffusion for indoor single-view material es- timation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5198– 5208, 2024. 2

    2024

  32. [32]

    Lightness and retinex theory.Journal of the Optical society of America, 61(1):1– 11, 1971

    Edwin H Land and John J McCann. Lightness and retinex theory.Journal of the Optical society of America, 61(1):1– 11, 1971. 2

    1971

  33. [33]

    A path towards autonomous machine intelli- gence version 0.9

    Yann LeCun. A path towards autonomous machine intelli- gence version 0.9. 2, 2022-06-27.Open Review, 62(1):1–62,

    2022

  34. [34]

    Materials for masses: Svbrdf acquisition with a single mobile phone image

    Zhengqin Li, Kalyan Sunkavalli, and Manmohan Chan- draker. Materials for masses: Svbrdf acquisition with a single mobile phone image. InProceedings of the Euro- pean Conference on Computer Vision (ECCV), pages 72–87,

  35. [35]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. 6

    Pith/arXiv arXiv 2017

  36. [36]

    Materia: Single image high- resolution material capture in the wild

    Rosalie Martin, Arthur Roullier, Romain Rouffet, Adrien Kaiser, and Tamy Boubekeur. Materia: Single image high- resolution material capture in the wild. InComputer Graph- ics Forum, pages 163–177. Wiley Online Library, 2022. 2, 3

    2022

  37. [37]

    Unsupervised learning of visual representations by solving jigsaw puzzles

    Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. InEuro- pean conference on computer vision, pages 69–84. Springer,

  38. [38]

    Repre- sentation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Repre- sentation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018. 6

    Pith/arXiv arXiv 2018

  39. [39]

    DINOv2: Learning robust visual features without supervi- sion.arXiv preprint arXiv:2304.07193, 2023

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervi- sion.arXiv preprint arXiv:2304.07193, 2023. 2, 3, 7

    Pith/arXiv arXiv 2023

  40. [40]

    Context encoders: Feature learning by inpainting

    Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 2536–2544, 2016. 3

    2016

  41. [41]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, pages 8748–8763. PMLR, 2021. 6

    2021

  42. [42]

    Textile: A differentiable metric for texture tileability

    Carlos Rodriguez-Pardo, Dan Casas, Elena Garces, and Jorge Lopez-Moreno. Textile: A differentiable metric for texture tileability. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 4439–4449, 2024. 3

    2024

  43. [43]

    Silhouettes: a graphical aid to the inter- pretation and validation of cluster analysis.Journal of com- putational and applied mathematics, 20:53–65, 1987

    Peter J Rousseeuw. Silhouettes: a graphical aid to the inter- pretation and validation of cluster analysis.Journal of com- putational and applied mathematics, 20:53–65, 1987. 8

    1987

  44. [44]

    Sfsnet: Learning shape, reflectance and illuminance of facesin the wild’

    Soumyadip Sengupta, Angjoo Kanazawa, Carlos D Castillo, and David W Jacobs. Sfsnet: Learning shape, reflectance and illuminance of facesin the wild’. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 6296–6305, 2018. 2

    2018

  45. [45]

    The effect of shape and illumination on material perception: model and applications.ACM Trans

    Ana Serrano, Bin Chen, Chao Wang, Michal Piovarci, Hans- Peter Seidel, Piotr Didyk, and Karol Myszkowski. The effect of shape and illumination on material perception: model and applications.ACM Trans. on Graph., 40(4), 2021. 4

    2021

  46. [46]

    Materialistic: Se- lecting similar materials in images.ACM Transactions on Graphics, 42(4), 2023

    Prafull Sharma, Julien Philip, Micha ¨el Gharbi, Bill Freeman, Fredo Durand, and Valentin Deschaintre. Materialistic: Se- lecting similar materials in images.ACM Transactions on Graphics, 42(4), 2023. 2, 3

    2023

  47. [47]

    Neural face editing with intrinsic image disentangling

    Zhixin Shu, Ersin Yumer, Sunil Hadap, Kalyan Sunkavalli, Eli Shechtman, and Dimitris Samaras. Neural face editing with intrinsic image disentangling. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 5541–5550, 2017. 2

    2017

  48. [48]

    Dinov3.arXiv preprint arXiv:2508.10104, 2025

    Oriane Sim ´eoni, Huy V V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha ¨el Ramamonjisoa, et al. Dinov3.arXiv preprint arXiv:2508.10104, 2025. 2, 3, 5, 6, 7

    Pith/arXiv arXiv 2025

  49. [49]

    Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063,

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063,

  50. [50]

    Trowbridge and K

    S. Trowbridge and K. P. Reitz. Average irregularity repre- sentation of a rough ray reflection.Journal of the Optical Society of America, 65(5):531–536, 1975. 4

    1975

  51. [51]

    Matsynth: A modern pbr materials dataset

    Giuseppe Vecchio and Valentin Deschaintre. Matsynth: A modern pbr materials dataset. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22109–22118, 2024. 2, 3

    2024

  52. [52]

    Surfacenet: Adversarial svbrdf estimation from a sin- gle image

    Giuseppe Vecchio, Simone Palazzo, and Concetto Spamp- inato. Surfacenet: Adversarial svbrdf estimation from a sin- gle image. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 12840–12848, 2021. 2

    2021

  53. [53]

    Controlmat: A controlled generative approach to material capture.ACM Transactions on Graphics, 43(5): 1–17, 2024

    Giuseppe Vecchio, Rosalie Martin, Arthur Roullier, Adrien Kaiser, Romain Rouffet, Valentin Deschaintre, and Tamy Boubekeur. Controlmat: A controlled generative approach to material capture.ACM Transactions on Graphics, 43(5): 1–17, 2024. 2

    2024

  54. [54]

    Terrain traversability prediction through self- supervised learning and unsupervised domain adaptation on synthetic data.Autonomous Robots, 48(2):4, 2024

    Giuseppe Vecchio, Simone Palazzo, Dario C Guastella, Daniela Giordano, Giovanni Muscato, and Concetto Spamp- inato. Terrain traversability prediction through self- supervised learning and unsupervised domain adaptation on synthetic data.Autonomous Robots, 48(2):4, 2024. 2 10

    2024

  55. [55]

    Microfacet models for refraction through rough surfaces

    Bruce Walter, Stephen R Marschner, Hongsong Li, and Ken- neth E Torrance. Microfacet models for refraction through rough surfaces. InProceedings of the 18th Eurographics conference on Rendering Techniques, pages 195–206, 2007. 4

    2007

  56. [56]

    RGB↔X: Image decomposition and synthe- sis 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 synthe- sis using material- and lighting-aware diffusion models. In Special Interest Group on Computer Graphics and Interac- tive Techniques Conference Conference Papers, page 1–11. ACM, 2024. 2

    2024

  57. [57]

    Colorful image colorization

    Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. InEuropean conference on computer vision, pages 649–666. Springer, 2016. 3

    2016

  58. [58]

    ibot: Image bert pre-training with online tokenizer.arXiv preprint arXiv:2111.07832,

    Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer.arXiv preprint arXiv:2111.07832,

    Pith/arXiv arXiv

  59. [59]

    Adversarial single-image svbrdf estimation with hybrid training

    Xilong Zhou and Nima Khademi Kalantari. Adversarial single-image svbrdf estimation with hybrid training. InCom- puter Graphics Forum, pages 315–325. Wiley Online Li- brary, 2021. 2 11

    2021

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.