Pith. sign in

REVIEW 3 major objections 4 minor 30 references

SAIL: Self-supervised Albedo Estimation from Real Images with a Latent Diffusion Model

T0 review · 3 major / 4 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 The paper claims that a latent diffusion model fine-tuned for unconditioned relighting can estimate stable albedo from a single real image using only unlabeled multi-illumination data.

desk verdict A sensible self-supervised latent-decomposition idea, but the headline benchmark was used to design the regularizer, so the outperformance claim does not hold as stated. read the letter →

arxiv 2505.19751 v2 pith:DAJISUN5 submitted 2025-05-26 cs.CV

Hala Djeghim , Nathan Piasco , Luis Roldão , Moussab Bennehar , Dzmitry Tsishkou , Céline Loscos , Désiré Sidibé This is my paper
classification cs.CV
keywords intrinsicimagedecompositionalbedoestimationself-supervisedlearninglatentdiffusionmodelrelightingmulti-illuminationdatasetslighting-invariantrepresentation
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

SAIL sets out to show that intrinsic albedo maps (the base color of surfaces independent of lighting) can be estimated from single-view real-world images without albedo labels. Its training signal is a surrogate task: the paper fine-tunes a pretrained latent diffusion model to relight a scene unconditionally, forcing the model to separate what stays the same across lighting changes from what changes. The decomposition is done entirely in latent space, with an image latent written as the sum of an albedo latent and a lighting latent, and several regularizers keep the albedo latent stable across illuminations. If the claim holds, albedo estimators could be trained from the abundant unlabeled multi-illumination imagery available online instead of relying on synthetic data or expensive ground truth.

What carries the argument

The load-bearing object is an additive intrinsic decomposition carried out in the latent space of a frozen VAE: $z_i = z_A + z_i^E$, where $z_A$ is the lighting-invariant albedo latent and $z_i^E$ is the latent component carrying illumination effects. A fine-tuned latent diffusion model (a generative model trained to denoise latents of a frozen autoencoder) is repurposed so that, given a noisy latent and a conditioning latent, it outputs both components, and their sum is trained to reconstruct the relighted latent. Regularizers then do the work: $L_{\mathrm{albedo}}$ enforces equal predicted albedo latents across illuminations, $L_{\mathrm{invariant}}$ keeps $z_A$ close to the input latent, and $L_{\mathrm{reg}} = \max(0, z_i^E)$ suppresses positive values in the lighting latent, preventing collapse and keeping high-frequency detail out of the lighting component.

What would settle it

Take the exact training setup but remove the correction term whose sign was chosen from MIDIntrinsics, and evaluate on a multi-illumination dataset whose true albedo comes from a different source, such as Blender renderings with known albedo. If the consistency advantage over LatentIntrinsics largely disappears, the claim that unlabeled data alone drives the result is in doubt; if it remains, the correction term is not decisive.

Watch

Extended reading notes

Core claim

The central claim is that a latent diffusion model pretrained for image generation can be repurposed into an intrinsic-image decomposition network without any labeled albedo. Training the model to perform unconditioned scene relighting forces it to encode the lighting-invariant content of the scene in one additive latent component $z_A$ and the lighting-dependent residual in another, with $z = z_A + z_E$. Decoding $z_A$ with the frozen VAE decoder yields the albedo-like output. On the MIDIntrinsics consistency benchmark the paper reports PSNR 17.00 and SSIM 0.63, and it argues that this stability under changing illumination is what distinguishes SAIL from earlier self-supervised albedo methods.

Load-bearing premise

The load-bearing premise is that the ground-truth albedo data used to decide one of the method's correction terms is also an unbiased test of its success; since the same MIDIntrinsics data appears on both sides, the headline consistency gain is not an independent check.

Editorial extensions

If this is right

  • Albedo maps predicted from a single real image become stable across illumination changes, which makes virtual relighting with arbitrary environment maps a direct downstream application.
  • Because training needs only unlabeled multi-illumination image pairs, the method can in principle be scaled with internet time-lapse and multi-illumination footage rather than annotated datasets.
  • Latent-space decomposition inherits the generative prior of a pretrained diffusion model, which is what the paper credits for removing reflections on glossy surfaces where earlier self-supervised methods fail.
  • On the MIDIntrinsics consistency benchmark, the paper reports PSNR 17.00 and SSIM 0.63, above both supervised diffusion baselines and the latent relighting baseline LatentIntrinsics.

Reading between the lines

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

  • If the additive latent split is sound, the same machinery could be pointed at other lighting-dependent factors—shadows, specular highlights, or even weather conditions—by choosing a different invariant target than albedo.
  • A version of SAIL that estimates the lighting-latent distribution from unlabeled pairs alone, rather than from MIDIntrinsics ground truth, would be needed to know how much of the success is genuinely self-supervised.
  • The reported 10-sample averaging at inference means part of the consistency gain may come from variance reduction; single-sample evaluation would isolate the contribution of the decomposition itself.
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

3 major / 4 minor

Summary. The paper proposes SAIL, a self-supervised method for albedo estimation from single real-world images. The approach fine-tunes a pre-trained latent diffusion model for unconditioned scene relighting, decomposes the input latent into an additive albedo latent and a lighting latent, and introduces several regularizers, including a non-positivity penalty on the lighting latent. The model is trained on unlabeled multi-illumination datasets and evaluated on MIDIntrinsics, IIW, and MAW. The authors claim that SAIL outperforms both supervised and self-supervised baselines in albedo consistency and robustness.

Significance. If substantiated, the contribution is significant: SAIL would demonstrate that a diffusion prior can be repurposed for self-supervised albedo estimation without any labeled data, using a latent-space intrinsic decomposition that is more flexible than image-space shading. The qualitative results and the use of online multi-illumination data are promising, and the paper includes a useful ablation study. However, the quantitative support for the central claim is currently undermined by the circular use of MIDIntrinsics for both regularizer design and evaluation, and by the in-the-wild metrics, which do not support the claimed superiority over baselines.

major comments (3)
  1. [Section 3.3, Eq. (9) and Appendix A] The non-positivity prior Lreg = max(0, zE_i) is designed by analyzing the distribution of the predicted lighting latent under full supervision on MIDIntrinsics ground-truth albedo, as described in Appendix A. The same MIDIntrinsics dataset is then used as the headline evaluation in Table 2. Since Table 4 shows that removing Lreg together with Linvariant degrades PSNR from 18.15 to 15.26 on the Toy Dataset, this regularizer is load-bearing. Consequently, the reported consistency advantage on MIDIntrinsics is not an independent test of the method; it is partly forced by a constraint fitted to the benchmark. The authors should derive the prior from a held-out split or a different dataset and re-report Table 2 with that clean protocol.
  2. [Section 4.2, Table 3 vs. abstract and Introduction] The abstract and Introduction claim that SAIL "outperforms both supervised and self-supervised baselines," but Table 3 does not support this. On IIW, SAIL's WHDR is 34% versus 21% for RGB→X; on MAW, it is 33% versus 22% for RGB→X. The paper's own text for Table 3 only claims "comparable results," which contradicts the stronger phrasing elsewhere. The central outperformance claim should be qualified to consistency-specific evaluation, or the authors should provide a detailed comparison that explains why the WHDR gap is acceptable.
  3. [Section 4.1 Datasets and Section 4.2] The training set is the MIT Multi-Illumination dataset, and the consistency evaluation uses MIDIntrinsics, which is an extension of the same dataset. The paper does not state whether the 30 MIDIntrinsics test scenes are disjoint from the 985 training scenes. If the scene-level split is not clean, the evaluation is further compromised. Please specify the exact split and confirm that no test scene appears during training, including when running the supervised experiment in Appendix A for designing Lreg.
minor comments (4)
  1. [Introduction and Figure 3 caption] The phrase "lightning de-correlated components" should read "lighting-decorrelated components"; the same typo appears in the Figure 3 caption where "lightning" should be "lighting."
  2. [Figure 3 caption] The caption says "we constraint the predicted latent albedo extracted from these sources latents"; this should be "we constrain the predicted latent albedo extracted from these source latents."
  3. [Section 4.2, Table 2] Table 2 reports mean PSNR and SSIM over 30 test scenes but gives no standard deviations or per-scene statistics; please add these and clarify the evaluation protocol, for example whether the 10-sample averaging is performed with fixed seeds.
  4. [Algorithm 1 and Section 4.1] The inference algorithm does not mention how classifier-free guidance with scale 1.5 is incorporated, although it is stated in Section 4.1; please align the algorithm pseudocode with the implementation details.

Circularity Check

1 steps flagged · score 7.0 of 10

Headline consistency claim rests on MIDIntrinsics, whose ground truth was used to design the Lreg regularizer; the Table 2 outperformance is not an independent test.

  1. fitted input called prediction [Section 3.3 (Eq. 9), Appendix A, and Section 4.2 / Table 2]
    "To estimate this distribution, we run an experiment in a fully supervised setup using ground truth albedo from the MIDIntrinsics dataset [4] and analyze the value of predicted ˆzE_i ... We found that the distribution of ˆzE_i is largely shifted toward negative values, hence we add a simple yet effective constraint on the model prediction: Lreg = max(0, ˆzE_i). ... We report in Tab. 2 the consistency evaluation results on the MIDIntrinsics dataset [4]. SAIL outperforms both supervised and self-supervised baselines."

    The regularizer Lreg (Eq. 9) is derived from a fully supervised experiment on MIDIntrinsics ground-truth albedo (Appendix A), and the same MIDIntrinsics ground truth is used as the evaluation target in Table 2 for the paper's headline superiority claim. Ablation Table 4 shows the regularization term is load-bearing: removing it and Linvariant drops PSNR from 18.15 to 15.26 on the synthetic Toy Dataset. Thus the Table 2 numbers do not independently validate self-supervised albedo estimation; they partly measure how well a benchmark-informed design constraint performs on the same benchmark.

full rationale

SAIL's core training objective (relighting surrogate, latent additive decomposition, consistency losses) is self-contained and not circular. However, the central quantitative claim of outperformance is made on MIDIntrinsics (Table 2), and the Lreg regularizer was designed using MIDIntrinsics ground-truth albedo in a supervised analysis (Section 3.3 and Appendix A). Because Lreg is load-bearing (Table 4), the benchmark evaluation is contaminated: the model is shaped by the test dataset's labels before being scored on it. The other public benchmarks (IIW/MAW, Table 3) do not support the 'outperforms' claim; the paper only claims SAIL achieves comparable results there. Therefore the headline consistency advantage reduces, at least in part, to a benchmark-informed design choice rather than an independent self-supervised result. This is a partial but substantive circularity in evaluation, not in the method's core formulation.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method rests on one ad hoc modeling choice (additive latent decomposition) and three domain assumptions inherited from prior work (relighting surrogacy, Retinex, VAE latent linearity). The free parameters are the loss weight lambda and the non-positivity prior, the latter of which was tuned on the evaluation benchmark.

free parameters (2)
  • Lambda (loss balancing factor) = 0.5
    Set by hand for all experiments (Section 3.4). It weights the regularization losses Linvariant and Lreg.
  • Non-positivity prior on zE (Lreg) = max(0, zE_i)
    Designed from a supervised analysis of MIDIntrinsics ground truth (Section 3.3, Appendix A); the same dataset is used for the headline evaluation in Table 2, so it acts as a test-set-derived constraint.
assumptions (4)
  • ad hoc to paper Latent additive decomposition: z_i = zA + zE_i
    Introduced in Eq. 3, this is the core modeling assumption without physical derivation. It is not obvious that the frozen VAE latent space is linearly decomposable into albedo and lighting.
  • domain assumption Relighting surrogate objective yields albedo
    Section 3.1 assumes that training a diffusion model to relight between illuminations forces the network to isolate lighting-invariant albedo. This is the key learning hypothesis.
  • domain assumption Retinex assumption: albedo is more similar to the image than shading
    Used to justify Linvariant (Eq. 8) in Section 3.3. This classical theory is adopted without validation for latent-space representations.
  • domain assumption The frozen Stable Diffusion VAE latent space supports additive decomposition
    Section 3.2 relies on the expressiveness of the Stable Diffusion latent space to host the additive decomposition, but no analysis is provided that this space is physically or linearly structured in the required way.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAIL: Self-supervised Albedo Estimation from Real Images with a Latent Diffusion Model." pith.science (2026). https://pith.science/paper/DAJISUN5

@misc{pith2026250519751,
  author       = {Pith},
  title        = {Pith review of: SAIL: Self-supervised Albedo Estimation from Real Images with a Latent Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DAJISUN5}},
  note         = {Machine review of arXiv:2505.19751}
}
read the original abstract

Intrinsic image decomposition aims at separating an image into its underlying albedo and shading components, isolating the base color from lighting effects to enable downstream applications such as virtual relighting and scene editing. Despite the rise and success of learning-based approaches, intrinsic image decomposition from real-world images remains a significant challenging task due to the scarcity of labeled ground-truth data. Most existing solutions rely on synthetic data as supervised setups, limiting their ability to generalize to real-world scenes. Self-supervised methods, on the other hand, often produce albedo maps that contain reflections and lack consistency under different lighting conditions. To address this, we propose SAIL, an approach designed to estimate albedo-like representations from single-view real-world images. We repurpose the prior knowledge of a latent diffusion model for unconditioned scene relighting as a surrogate objective for albedo estimation. To extract the albedo, we introduce a novel intrinsic image decomposition fully formulated in the latent space. To guide the training of our latent diffusion model, we introduce regularization terms that constrain both the lighting-dependent and independent components of our latent image decomposition. SAIL predicts stable albedo under varying lighting conditions and generalizes to multiple scenes, using only unlabeled multi-illumination data available online.

Figures

Figures reproduced from arXiv: 2505.19751 by the authors.

Figure 1
Figure 1. Self-Supervised Albedo Estimation from Real Images – From a single image under real-world lighting conditions, SAIL extracts high-fidelity albedo by repurposing and finetuning a pretrained latent diffusion model (left). The estimated albedo enables downstream tasks such as single-image virtual relighting, demonstrated using Blender [6] with different environment maps (right). Abstract Intrinsic image decomposition a… view at source ↗
Figure 2
Figure 2. SAIL overview – Given a single input image, encoded into the latent space using the frozen pre-trained VAE encoder, SAIL estimates an albedo representation in latent space, which, when decoded, produces an albedo image without any lighting effects. a shading component, which captures illumination effects. Early solutions [10] were inspired by Retinex Theory [15], where large intensity gradients in an image are attri… view at source ↗
Figure 3
Figure 3. SAIL training – SAIL performs image intrinsic decomposition (zˆ A i , zˆ E i ) conditioned on a source latent zi , through a surrogate objective of image relighting (Lrelight and Lconsistency). Considering multiple illuminations i, j of the same scene, we constraint the predicted latent albedo extracted from these sources latents to be identical (Lalbedo). characterized by its full representation in the latent space… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: We qualitatively compare the predicted albedos on the MITDataset [ [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: We qualitatively compare the predicted albedos on the BigTime [ [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: We qualitatively compare the predicted albedos on the MAW dataset [ [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Analysis of of predicted latent distribution [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: We show results of virtual relighting with Blender [ [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: We show unconditioned relighting results predicted by SAIL at different inferences. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: We qualitatively compare the predicted albedos on the IIW dataset [ [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 25 canonical work pages

  1. [1]

    Recovering intrinsic scene characteris- tics

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

    1978

  2. [2]

    Intrinsic images in the wild

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

    work page 2014

  3. [3]

    Anand Bhattad, James Soole, and D.A. Forsyth. Stylitgan: Image-based relighting via la- tent control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4231–4240, June 2024

    work page 2024

  4. [4]

    Intrinsic image decomposition via ordinal shading

    Chris Careaga and Ya˘gız Aksoy. Intrinsic image decomposition via ordinal shading. ACM Trans. Graph., 2023

    work page 2023

  5. [5]

    Mair: multi-view attention inverse rendering with 3d spatially-varying lighting estimation

    JunYong Choi, SeokYeong Lee, Haesol Park, Seung-Won Jung, Ig-Jae Kim, and Junghyun Cho. Mair: multi-view attention inverse rendering with 3d spatially-varying lighting estimation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 8392–8401, 2023

    work page 2023

  6. [6]

    Blender - a 3D modelling and rendering package

    Blender Online Community. Blender - a 3D modelling and rendering package . Blender Foundation, 2018

    work page 2018

  7. [7]

    Generative models: What do they know? do they know things? let’s find out! arXiv preprint arXiv:2311.17137, 2023

    Xiaodan Du, Nicholas Kolkin, Greg Shakhnarovich, and Anand Bhattad. Generative models: What do they know? do they know things? let’s find out! arXiv preprint arXiv:2311.17137, 2023

    arXiv 2023

  8. [8]

    Unsupervised learning of compositional energy concepts

    Yilun Du, Shuang Li, Yash Sharma, Josh Tenenbaum, and Igor Mordatch. Unsupervised learning of compositional energy concepts. Advances in Neural Information Processing Systems, 34:15608–15620, 2021

    work page 2021

Show all 30 references
  1. [9]

    Intrinsic image decomposition using paradigms

    David Forsyth and Jason J Rock. Intrinsic image decomposition using paradigms. IEEE transactions on pattern analysis and machine intelligence , 44(11):7624–7637, 2021

    2021

  2. [10]

    A survey on intrinsic images: Delving deep into lambert and beyond

    Elena Garces, Carlos Rodriguez-Pardo, Dan Casas, and Jorge Lopez-Moreno. A survey on intrinsic images: Delving deep into lambert and beyond. International Journal of Computer Vision, 130(3):836–868, 2022

    2022

  3. [11]

    Ground truth dataset and baseline evaluations for intrinsic image algorithms

    Roger Grosse, Micah K Johnson, Edward H Adelson, and William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In 2009 IEEE 12th International Conference on Computer Vision, pages 2335–2342, 2009

    2009

  4. [12]

    Classifier-free diffusion guidance, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022

    2022

  5. [13]

    Intrinsic image diffusion for indoor single-view material estimation

    Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. Intrinsic image diffusion for indoor single-view material estimation. Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 5198–5208, 2024

    2024

  6. [14]

    Intrinsic decomposition of image sequences from local temporal variations

    Pierre-Yves Laffont and Jean-Charles Bazin. Intrinsic decomposition of image sequences from local temporal variations. In Proceedings of the IEEE international conference on computer vision, pages 433–441, 2015

    2015

  7. [15]

    Lightness and retinex theory

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

    1971

  8. [16]

    Learning intrinsic image decomposition from watching the world

    Zhengqi Li and Noah Snavely. Learning intrinsic image decomposition from watching the world. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 9039–9048, 2018

    2018

  9. [17]

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

    Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern ...

    2020

  10. [18]

    A review of intrinsic image decomposition

    Siyuan Liu, Xiaoyue Jiang, Letian Liu, Zhaoqiang Xia, Sihang Dang, and Xiaoyi Feng. A review of intrinsic image decomposition. In 2024 3rd International Conference on Image Processing and Media Computing (ICIPMC) , pages 254–261, 2024

    2024

  11. [19]

    Interactive Virtual Relighting and Remodeling of Real Scenes

    Céline Loscos, Marie-Claude Frasson, George Drettakis, Bruce Walter, Xavier Granier, and Pierre Poulin. Interactive Virtual Relighting and Remodeling of Real Scenes. In Rendering Techniques ’99, pages 329–340, 1999

    1999

  12. [20]

    Single image intrinsic decomposition without a single intrinsic image

    Wei-Chiu Ma, Hang Chu, Bolei Zhou, Raquel Urtasun, and Antonio Torralba. Single image intrinsic decomposition without a single intrinsic image. In Proceedings of the European conference on computer vision (ECCV) , pages 201–217, 2018

    2018

  13. [21]

    A multi-illumination dataset of indoor object appearance

    Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A multi-illumination dataset of indoor object appearance. In 2019 IEEE International Conference on Computer Vision (ICCV), 2019

    2019

  14. [22]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 10684–10695, 2022

    2022

  15. [23]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR) , 2021

    2021

  16. [24]

    Tenenbaum, and Yilun Du

    Jocelin Su, Nan Liu, Yanbo Wang, Joshua B. Tenenbaum, and Yilun Du. Compositional image decomposition with diffusion models. In Proceedings of the 41st International Conference on Machine Learning, pages 46823–46842, 2024

    2024

  17. [25]

    Deriving intrinsic images from image sequences

    Yair Weiss. Deriving intrinsic images from image sequences. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001 , volume 2, pages 68–75, 2001

    2001

  18. [26]

    Measured albedo in the wild: Filling the gap in intrinsics evaluation

    Jiaye Wu, Sanjoy Chowdhury, Hariharmano Shanmugaraja, David Jacobs, and Soumyadip Sengupta. Measured albedo in the wild: Filling the gap in intrinsics evaluation. In 2023 IEEE International Conference on Computational Photography (ICCP) , pages 1–12, 2023

    2023

  19. [27]

    Rgb ↔x: Image decomposition and synthesis using material- and lighting-aware diffusion models

    Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hašan. Rgb ↔x: Image decomposition and synthesis using material- and lighting-aware diffusion models. In Special Interest Group on Computer Graphics and Inter...

    2024

  20. [28]

    Latent intrinsics emerge from training to relight

    Xiao Zhang, William Gao, Seemandhar Jain, Michael Maire, David Forsyth, and Anand Bhattad. Latent intrinsics emerge from training to relight. Advances in Neural Information Processing Systems, 37:96775–96796, 2024

    2024

  21. [29]

    I2-sdf: Intrinsic indoor scene reconstruction and editing via raytracing in neural sdfs

    Jingsen Zhu, Yuchi Huo, Qi Ye, Fujun Luan, Jifan Li, Dianbing Xi, Lisha Wang, Rui Tang, Wei Hua, Hujun Bao, et al. I2-sdf: Intrinsic indoor scene reconstruction and editing via raytracing in neural sdfs. In Proceedings of the IEEE/CVF conference on computer vision and pattern ...

    2023

  22. [30]

    Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing

    Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers, pages 1–8...

    2022

Pith tools

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