Pith. sign in

REVIEW 3 major objections 7 minor 33 references

SPC to 3D: Novel View Synthesis from Binary SPC via I2I translation

T0 review · 3 major / 7 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 Two-stage translation turns binary single-photon images into color 3D views.

desk verdict A plausible but under-supported benchmark: two-stage I2I+NeRF/3DGS is new for binary SPC, but the uncalibrated simulator and missing validation leave the ranking unproven. read the letter →

arxiv 2506.06890 v1 pith:BS7KS2WB submitted 2025-06-07 eess.IV cs.CVeess.SP

Sumit Sharma , Gopi Raju Matta , Kaushik Mitra This is my paper
classification eess.IVcs.CVeess.SP
keywords single-photoncamerasSPADbinaryimagetranslationPix2PixHDNeuralRadianceFields3DGaussianSplattingnovelviewsynthesisimage-to-image
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

Single-photon cameras capture light with extreme sensitivity but output only binary images, stripping away the texture and color that NeRF and 3D Gaussian Splatting need. This paper argues that a two-stage pipeline fixes the problem: first translate the binary SPC frames into plausible RGB images with an image-to-image model, then run NeRF or 3D Gaussian Splatting on those RGB images to synthesize novel views. Across an eight-way comparison on simulated SPC data, the paper reports that Pix2PixHD followed by NeRF or 3DGS gives the highest PSNR, SSIM, and LPIPS scores, with Pix2PixHD + NeRF reaching 22.70 PSNR and 0.6843 SSIM. The paper also contributes a physics-inspired SPC simulator, based on Poisson photon arrivals and SPAD dead time, to produce paired binary/RGB training data.

What carries the argument

The load-bearing component is the two-stage modular pipeline. Stage one is image-to-image translation: Pix2PixHD, whose multi-scale generator and discriminator with feature-matching loss convert binary SPC inputs into RGB images. Stage two is radiance-field reconstruction: NeRF (a fully connected network mapping position and viewing direction to color and density) or 3D Gaussian Splatting (a rasterized collection of 3D Gaussians) renders novel views from the translated RGB frames. Supporting this is the SPC simulator: it treats photon arrivals as a Poisson renewal process with a dead-time model, using equations for the mean and variance of detected photon counts (with quantum efficiency 0.45 and dead time 150 ns), then thresholds counts to binary detections per color channel.

What would settle it

Train and evaluate the same pipelines on real SPC recordings with ground-truth RGB captures from the same scene geometry; if Pix2PixHD + NeRF/3DGS no longer beats the denoising baselines or direct reconstruction on PSNR/SSIM/LPIPS, the paper's central claim is refuted. A less expensive check is to compare the photon-count statistics of the simulator against real sensor histograms under controlled flux.

Watch

Extended reading notes

Core claim

The central claim is that binary single-photon-camera images, which look like black-and-white noise, can be turned into high-quality color novel views by decoupling the problem into two stages rather than forcing 3D reconstruction to operate directly on the binary data. The first stage uses Pix2PixHD, a conditional GAN with a feature-matching adversarial loss, to hallucinate color and texture from the binary input; the second stage feeds those RGB images to NeRF or 3D Gaussian Splatting. The paper reports that this combination outperforms direct single-stage NeRF/3DGS, denoise-then-reconstruct baselines (Restormer and DiffBIR), and other translation models (Pix2Pix, CycleGAN, LookatMe), on simulated scenes from the IBRNet-Collected-2 dataset.

Load-bearing premise

The entire evaluation rests on the simulated SPC images faithfully matching what a real single-photon camera would output; if the simulator's noise, dead-time, and quantum-efficiency assumptions are off, the reported ranking of pipelines may not carry over to actual hardware.

Editorial extensions

If this is right

  • Pix2PixHD + NeRF reaches PSNR 22.70 and SSIM 0.6843 on simulated SPC novel-view synthesis, the best of the eight compared combinations.
  • Pix2PixHD + 3D Gaussian Splatting attains the best LPIPS (0.4251), meaning perceptual quality is strongest when the second stage is 3DGS.
  • Direct single-stage reconstruction (NeRF or 3DGS on raw binary SPC images) produces washed-out or noisy renders and lower scores, showing that translation before reconstruction is necessary.
  • Denoising binary SPC images with Restormer or DiffBIR before the 3D stage does not match the translation-first pipeline, producing lighter colors and added noise.
  • The two-stage design is modular, so each stage can be swapped and optimized independently.

Reading between the lines

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

  • If the simulator's Poisson/dead-time model closely matches real SPC sensors, the same two-stage pipeline should transfer to hardware; the paper lists real-sensor validation as future work, so this remains a testable prediction rather than a demonstrated result.
  • The success of a GAN-based translator suggests that other powerful conditional generation models could further close the gap with ground-truth RGB, potentially improving the downstream 3D stage without changing the reconstruction module.
  • The modularity invites end-to-end joint training of translation and reconstruction, which the paper names as future work; a single loss could encourage translation outputs that are optimized for geometric consistency, not just visual similarity.
  • Because the simulator thresholds photon counts at zero, extremely low-flux regions become uniformly black bits; a real sensor would exhibit thermal noise and non-uniform quantum efficiency, which could change which translation model wins.
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 / 7 minor

Summary. The paper proposes a modular two-stage pipeline for novel view synthesis from binary Single Photon Camera (SPC) images. Stage 1 translates simulated binary SPC images into RGB-like images using image-to-image translation models (Pix2PixHD, Pix2Pix, CycleGAN, LookatMe, Restormer, DiffBIR), and Stage 2 feeds the translated RGB images into NeRF or 3D Gaussian Splatting for novel view rendering. A physics-inspired SPC simulator based on Poisson photon arrivals and SPAD dead time is used to generate paired training and evaluation data. Quantitative results on simulated IBRNet scenes (Table 1) and qualitative comparisons (Fig. 2) lead the authors to claim that Pix2PixHD + NeRF/3DGS outperforms other combinations in perceptual quality and geometric consistency, with single-stage NeRF/3DGS baselines performing worse (Table 2).

Significance. If established, the result is practically relevant: it offers a modular recipe for turning a severely quantized, binary sensing modality into a form usable by standard radiance-field pipelines, and it systematically compares several I2I backbones for this task. The strengths of the paper are its clearly stated modular framework, the inclusion of a statistically motivated SPC simulator, and the breadth of the combination table (eight two-stage pipelines plus single-stage ablations). However, the central quantitative claim rests entirely on a synthetic benchmark generated by the authors' own simulator, with no fixed, reported physical operating point, no error bars or per-scene breakdowns, and no validation against real SPC hardware. The claimed ranking is therefore plausible but not yet demonstrated to be robust, which is what limits the current significance.

major comments (3)
  1. [§3.1, Eq. (1)] The SPC simulator's physical operating point is not fixed. Equation (1) shows that the mean detected count E[N_T^SPC] = q_SPAD * phi * T / (1 + q_SPAD * phi * tau_d) depends on the product q*phi*T, yet the text states only that 'photon flux phi represents the pixel intensity values' and does not report absolute flux units or exposure time T. The paper states that q_SPAD = 0.45 and tau_d = 150 ns are fixed, but T is left unspecified and the simulation procedure says T and q are 'adjusted' to ensure visibility. With T free, the same RGB image can be rendered as almost-all-zero or almost-all-one binary frames, which changes the difficulty of the I2I translation task dramatically. The headline ranking in Table 1 is therefore not shown to be robust even within the simulator, because no sweep over T or flux is reported. This is a load-bearing issue: the central claim 'Pix2PixHD + NeRF/3DGS outperforms other combinations' is only established for one uncalibrated operating point.
  2. [§4, Table 1] There is an internal inconsistency between the text and Table 1. Section 4 states that CycleGAN and LookatMe 'failed to successfully translate SPC binary images (3-channel)', yet Table 1 lists CycleGAN + NeRF with PSNR 16.49. Similarly, Restormer and DiffBIR are described as 'inefficient', yet they appear as rows in Table 1. The manuscript never explains how these rows were produced or whether the listed numbers represent a failure mode or a valid baseline. In addition, Table 1 reports no error bars, no seeds, no per-scene breakdown, and no statistical significance testing; the margins between Pix2PixHD+NeRF (PSNR 22.70) and Restormer+NeRF (PSNR 20.42), and between Pix2PixHD+3DGS (22.21) and DiffBIR+NeRF (19.35), could be within run-to-run variation. The paper should either remove the failed baselines from Table 1 or explain what the numbers mean, and should provide variance information.
  3. [§5, Conclusion] The evaluation is entirely synthetic: SPC images for both training and testing are generated by the authors' own simulator, and the conclusion explicitly defers real-sensor evaluation to future work. The central claim of practical relevance for binary SPC imagery is therefore not directly supported by the presented experiments. The simulator is grounded in a Poisson/dead-time model, but no comparison to real SPC measurements, no parameter-calibration procedure, and no sensitivity analysis are provided. A concrete test would be to evaluate the trained two-stage pipeline on real SPC captures (even a single scene) or, at minimum, to validate that the simulator's binary statistics match a real sensor at the chosen operating point. Without this, the reported ranking may be an artifact of the simulator's assumptions rather than a property of real SPC data.
minor comments (7)
  1. [Title] The title contains a typo: 'TRANSLA TION' should be 'TRANSLATION'.
  2. [§3.1, Eq. (1)-(2)] Equation (2) is not derived or cited precisely; the reference list includes Grimmett and Stirzaker for renewal processes and cites 'Ingle and Proakis' in the text, but the cited reference [21] is by Ingle, Velten, and Gupta, not Proakis. Please correct the attribution.
  3. [§3.1] The text uses inconsistent notation: the exposure time is introduced as 'T', then written as 'NT SP C', and q_SPAD appears in both the equations and text. Please unify the notation and define all variables at first use.
  4. [§4] The phrase 'approximately 8 feasible working combinations' is unclear: Table 1 contains exactly eight rows, but the text says CycleGAN failed and denoising was inefficient. Please clarify which combinations are considered 'working' and which are reported only as negative baselines.
  5. [§4, Fig. 2] The qualitative discussion relies heavily on visual inspection ('please zoom in'), but the figure is not available in the manuscript text. Please ensure the figure is legible at publication resolution and that the scene names and view indices are labeled.
  6. [§4.1] No training details are given for the I2I models (learning rate, batch size, number of epochs, resolution, train/test split). Without these, the comparison in Table 1 is not reproducible. At minimum, please add a supplementary table of hyperparameters and dataset splits.
  7. [General] The paper would benefit from a code/data release statement and from reporting the number of scenes used for evaluation. The claim that Pix2PixHD's feature-matching loss is 'pivotal' is plausible but is not directly ablated; an ablation removing the feature-matching loss would strengthen the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the two-stage pipeline is evaluated empirically against alternative combinations, with no prediction reducing to a fitted parameter or self-citation chain.

full rationale

The paper makes an empirical, comparative claim: Pix2PixHD followed by NeRF or 3DGS outperforms other I2I+3D combinations on simulated SPC inputs (Table 1). This claim is supported by a benchmark of eight pipelines; no parameter of the proposed method is fitted to the reported PSNR/SSIM/LPIPS values, and no output quantity is definitionally equal to an input quantity. The SPC simulator in Section 3.1 uses Eq. (1)-(2), which are attributed to external references [19] and [21] (Poisson statistics and dead-time renewal process), not to the authors' own prior work. Although reference [17] is a self-citation and the same simulator generates both training and test data, this creates a self-referential evaluation setting rather than a circular derivation: the ranking among methods is not forced by construction, and the physics model itself is cited to independent sources. The lack of validation on real SPC sensor data is a legitimate external-validity and correctness risk, and the internal inconsistency that CycleGAN is said to 'fail to translate' while Table 1 lists CycleGAN+NeRF is a reporting concern, but neither pattern matches the circularity definitions (self-definitional, fitted-input-as-prediction, load-bearing self-citation, imported uniqueness, ansatz-by-citation, or renaming). Therefore no circular step can be quoted and exhibited, and the paper receives a non-circular score.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The core approach depends on hand-set simulator constants and modeling assumptions; no new physical entities are introduced.

free parameters (4)
  • Quantum efficiency q_SPAD = 0.45
    Hand-set in the SPC simulator (Section 3.1) to control photon detection probability; affects all simulated binary images.
  • Dead time tau_d = 150 ns
    Fixed in the simulator (Section 3.1); determines mean and variance of detected photon counts in Equations 1 and 2.
  • Exposure time T = not specified
    Adjusted during simulation to ensure visibility of all regions (Section 3.1); exact values per dataset not reported.
  • Binary detection threshold = count > 0
    Simulated photon counts are thresholded at one or more photons to produce binary pixels (Section 3.1).
assumptions (5)
  • domain assumption Photon arrivals per pixel follow Poisson statistics.
    Invoked in Section 3.1 to compute mean and variance of detected photons.
  • standard math The detection process is a renewal process with dead time, with mean and variance given by Equations 1 and 2 from Ingle et al.
    Taken from prior literature [20, 21] and used without derivation.
  • domain assumption The SPC operates in passive mode and each detection triggers a fixed dead-time interval.
    Section 3.1, governs the simulator behavior.
  • domain assumption Input color image intensities can be treated as photon flux phi.
    Section 3.1, 'Once the input image is loaded, it is processed by extracting the photon flux Phi, which represents the pixel intensity values.'
  • ad hoc to paper Binary threshold at a detected count greater than zero preserves enough information for downstream colorization.
    Section 3.1, thresholding choice; central to why I2I can recover color.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SPC to 3D: Novel View Synthesis from Binary SPC via I2I translation." pith.science (2026). https://pith.science/paper/BS7KS2WB

@misc{pith2026250606890,
  author       = {Pith},
  title        = {Pith review of: SPC to 3D: Novel View Synthesis from Binary SPC via I2I translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BS7KS2WB}},
  note         = {Machine review of arXiv:2506.06890}
}
read the original abstract

Single Photon Avalanche Diodes (SPADs) represent a cutting-edge imaging technology, capable of detecting individual photons with remarkable timing precision. Building on this sensitivity, Single Photon Cameras (SPCs) enable image capture at exceptionally high speeds under both low and high illumination. Enabling 3D reconstruction and radiance field recovery from such SPC data holds significant promise. However, the binary nature of SPC images leads to severe information loss, particularly in texture and color, making traditional 3D synthesis techniques ineffective. To address this challenge, we propose a modular two-stage framework that converts binary SPC images into high-quality colorized novel views. The first stage performs image-to-image (I2I) translation using generative models such as Pix2PixHD, converting binary SPC inputs into plausible RGB representations. The second stage employs 3D scene reconstruction techniques like Neural Radiance Fields (NeRF) or Gaussian Splatting (3DGS) to generate novel views. We validate our two-stage pipeline (Pix2PixHD + Nerf/3DGS) through extensive qualitative and quantitative experiments, demonstrating significant improvements in perceptual quality and geometric consistency over the alternative baseline.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 28 canonical work pages

  1. [17]

    Photographic image synthesis with cascaded refinement networks,

    Qifeng Chen and Vladlen Koltun, “Photographic image synthesis with cascaded refinement networks,” inPro- ceedings of the IEEE international conference on com- puter vision, 2017, pp. 1511–1520

    work page 2017

  2. [1]

    INTRODUCTION Single-photon camera pixel possesses remarkable sensitivity, capable of detecting individual photons and recording their arrival times with picosecond precision. Single Photon Cam- eras (SPCs), built on SPAD arrays, can operate in extremely low-light environments and at high temporal resolutions, making them highly attractive for scientific, ...

    work page

  3. [2]

    RELA TED WORK 2.1. Image-to-Image translation Generative adversarial networks (GANs) [1] are designed to model the distribution of natural images by generating sam- ples that are indistinguishable from real-world images. They have enabled diverse applications, including image genera- tion [2, 3, 4]. Generative image models are broadly catego- rized into p...

    work page Pith review arXiv 2025

  4. [3]

    TWO STAGE APPROACH FOR NOVEL VIEW SYNTHESIS 3.1. Single Photon Camera Simulator To train various Image-to-Image translation networks, we constructed a paired dataset comprising Single Photon Cam- era (SPC) images and their corresponding RGB color images. Despite continuous advancements in SPC technology and its functionalities, their availability in the c...

    work page

  5. [4]

    EXPERIMENTS To train the Image-to-Image translation networks, we curated a comprehensive training dataset leveraging the NeRF-LLFF dataset [26]. To enrich the diversity of the training data, we applied a series of image augmentations, including zoom- ing, rotation, shearing, and flipping, which resulted in ap- proximately 15,000 augmented images. Subseque...

    work page

  6. [5]

    CONCLUSION In this paper, we presented a modular two-stage framework that transforms binary SPC images into color novel views. By integrating Pix2PixHD for image translation with NeRF or Gaussian Splatting for 3D reconstruction, our method demonstrates notable improvements in texture detail and per- ceptual quality over other combinations. The use of a fe...

    work page

  7. [6]

    Example-based super-resolution,

    William T Freeman, Thouis R Jones, and Egon C Pasz- tor, “Example-based super-resolution,”IEEE Computer graphics and Applications, vol. 22, no. 2, pp. 56–65, 2002

    work page 2002

  8. [7]

    Generative adversar- ial networks,

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio, “Generative adversar- ial networks,”Communications of the ACM, vol. 63, no. 11, pp. 139–144, 2020

    work page 2020

Show all 33 references
  1. [8]

    Wasserstein generative adversarial networks,

    Martin Arjovsky, Soumith Chintala, and L ´eon Bottou, “Wasserstein generative adversarial networks,” inInter- national conference on machine learning. PMLR, 2017, pp. 214–223

    2017

  2. [9]

    Unsupervised representation learn- ing with deep convolutional generative adversarial net- works,

    Alec Radford, “Unsupervised representation learn- ing with deep convolutional generative adversarial net- works,”arXiv preprint arXiv:1511.06434, 2015

    2015 arXiv

  3. [10]

    Energy-based generative adversarial net- work,

    Junbo Zhao, “Energy-based generative adversarial net- work,”arXiv preprint arXiv:1609.03126, 2016

    2016 arXiv

  4. [11]

    Texture syn- thesis by non-parametric sampling,

    Alexei A Efros and Thomas K Leung, “Texture syn- thesis by non-parametric sampling,” inProceedings of the seventh IEEE international conference on computer vision. IEEE, 1999, vol. 2, pp. 1033–1038

    1999

  5. [12]

    High-resolution image synthesis and semantic manipulation with con- ditional gans,

    Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro, “High-resolution image synthesis and semantic manipulation with con- ditional gans,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 8798–8807

    2018

  6. [13]

    Scene completion using millions of photographs,

    James Hays and Alexei A Efros, “Scene completion using millions of photographs,”ACM Transactions on graphics (TOG), vol. 26, no. 3, pp. 4–es, 2007

    2007

  7. [14]

    Image-to-image translation with conditional ad- versarial networks,

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros, “Image-to-image translation with conditional ad- versarial networks,” inProceedings of the IEEE confer- ence on computer vision and pattern recognition, 2017, pp. 1125–1134

    2017

  8. [15]

    Fully convolutional networks for semantic segmenta- tion,

    Jonathan Long, Evan Shelhamer, and Trevor Darrell, “Fully convolutional networks for semantic segmenta- tion,” inProceedings of the IEEE conference on com- puter vision and pattern recognition, 2015, pp. 3431– 3440

    2015

  9. [16]

    Holistically-nested edge detection,

    Saining Xie and Zhuowen Tu, “Holistically-nested edge detection,” inProceedings of the IEEE international conference on computer vision, 2015, pp. 1395–1403

    2015

  10. [18]

    Photon-starved scene inference using single photon cameras,

    Bhavya Goyal and Mohit Gupta, “Photon-starved scene inference using single photon cameras,” inProceedings of the IEEE/CVF International Conference on Com- puter Vision, 2021, pp. 2512–2521

    2021

  11. [19]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” inECCV, 2020

    2020

  12. [20]

    Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields,

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan, “Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, ...

    2021

  13. [21]

    Instant neural graphics primitives with a multiresolution hash encoding,

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexander Keller, “Instant neural graphics primitives with a multiresolution hash encoding,”TOG, 2022

    2022

  14. [22]

    3d gaussian splatting for real-time radiance field rendering,

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk¨uhler, and George Drettakis, “3d gaussian splatting for real-time radiance field rendering,”TOG, 2023

    2023

  15. [23]

    Transforming single photon camera images to color high dynamic range images,

    Sumit Sharma, Girish Rongali, and Kaushik Mi- tra, “Transforming single photon camera images to color high dynamic range images,”arXiv preprint arXiv:2412.12942, 2024

    2024 arXiv

  16. [24]

    However, this method resulted in ren- dered images that were lighter in color, lacked texture details, and exhibited increased noise

    and DiffBIR [25]. However, this method resulted in ren- dered images that were lighter in color, lacked texture details, and exhibited increased noise. For visual reference, see Fig. 2. 4.4. Ablations with Single Stage Approaches In Single Stage approaches, we utilized two dif...

  17. [25]

    Single-photon camera guided extreme dynamic range imaging,

    Yuhao Liu, Felipe Gutierrez-Barragan, Atul Ingle, Mo- hit Gupta, and Andreas Velten, “Single-photon camera guided extreme dynamic range imaging,” inProceed- ings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, 2022, pp. 1575–1585

    2022

  18. [26]

    Geoffrey Grimmett and David Stirzaker,Probability and random processes, Oxford university press, 2020

    2020

  19. [27]

    High flux passive imaging with single-photon sensors,

    Atul Ingle, Andreas Velten, and Mohit Gupta, “High flux passive imaging with single-photon sensors,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 6760–6769

    2019

  20. [28]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” inProceedings of the IEEE international conference on computer vi- sion, 2017, pp. 2223–2232

    2017

  21. [29]

    Look atme: the discriminator mean entropy needs attention,

    Edgardo Solano-Carrillo, Angel Bueno Rodriguez, Borja Carrillo-Perez, Yannik Steiniger, and Jannis Stoppe, “Look atme: the discriminator mean entropy needs attention,” inProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2023, pp. 787–796

    2023

  22. [30]

    Restormer: Efficient transformer for high- resolution image restoration,

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang, “Restormer: Efficient transformer for high- resolution image restoration,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5728–5739

    2022

  23. [31]

    Diffbir: Toward blind image restoration with generative diffusion prior,

    Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong, “Diffbir: Toward blind image restoration with generative diffusion prior,” inEuropean Conference on Computer Vision. Springer, 2025, pp. 430–448

    2025

  24. [32]

    Local light field fu- sion: Practical view synthesis with prescriptive sam- pling guidelines,

    Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz- Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar, “Local light field fu- sion: Practical view synthesis with prescriptive sam- pling guidelines,”ACM Transactions on Graphics (ToG), vol. 38, no. 4, pp....

    2019

  25. [33]

    Ibrnet: Learning multi-view image-based rendering,

    Howard Zhou, Jon Barron, Noah Snavely, Pratul Srini- vasan, Qianqian Wang, Ricardo Martin-Brualla, Tom Funkhouser, and Zhicheng Wang, “Ibrnet: Learning multi-view image-based rendering,” 2021

    2021

Pith tools

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