Pith. sign in

REVIEW 4 major objections 7 minor 55 references

CtrlNeRF: The Generative Neural Radiation Fields for the Controllable Synthesis of High-fidelity 3D-Aware Images

T0 review · 4 major / 7 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read CtrlNeRF shows that a single MLP with shared weights can represent multiple 3D scenes and control shape and color through label-embedded latent codes, keeping storage flat as scenes grow.

desk verdict A plausible GRAF extension for label-controlled multi-scene 3D generation whose central storage-quality claim outruns its evidence; worth a serious referee but needs a real baseline comparison. read the letter →

arxiv 2412.00754 v1 pith:RLRDVPOK submitted 2024-12-01 cs.CV cs.AI

Jian Liu , Zhen Yu This is my paper · ORCID
classification cs.CVcs.AI
keywords ImplicitRepresentationNovelViewSynthesisGenerativeAdversarialNetworkNeuralRadianceFieldControllableImageGeneration3D-AwareImagesLabelEmbeddingMulti-scene
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

CtrlNeRF sets out to show that a single MLP with shared weights can represent multiple 3D scenes and generate images whose shape follows a class label and whose appearance follows a color label. The method takes the generative radiance field (GRAF) and changes the input and output of its MLP: shape and appearance latent codes are multiplied by learned label embeddings, and the network outputs an array of densities and an array of colors indexed by class and style. A VGG-based discriminator is added to classify class and color, and the adversarial loss is augmented with class and style losses. If the claims hold, multi-scene 3D generation no longer needs one network per scene, and storage and training cost stay flat as scenes grow. The paper also reports that image quality degrades as the number and complexity of scenes grows, which it attributes to entanglement in the shared MLP.

What carries the argument

The load-bearing object is the label-embedded conditional radiance field: a GRAF-style MLP whose input is $\gamma(x)+z'_s$ and $\gamma(d)+z'_a$, where $z'_s$ and $z'_a$ are the latent codes multiplied by learned embeddings of the class and color labels, and whose output is a density array and a color array instead of a single density and color. The output arrays let one shared network expose separate slots for classes and styles, and the VGG-based discriminator supplies the classification signal that pushes those slots apart. Volume rendering of the arrays yields the 3D-consistent images, and linear interpolation between array entries yields the unseen features.

What would settle it

Render the density slot for one class while feeding the color embedding of another class into the same trained model; if the output geometry changes, the claimed independence of shape and appearance fails. Equivalently, if the FID-versus-class-count curve on the CARs dataset is flat rather than rising, the reported capacity degradation is not the real mechanism.

Watch

Extended reading notes

Core claim

CtrlNeRF's central claim is that a generative radiance field can be made controllable across multiple scenes without adding networks. By multiplying the shape code $z_s$ and appearance code $z_a$ by label-embedding vectors, and by changing the MLP output into a density array and a color array, a single weight-shared MLP learns to separate scenes by class and style; volume rendering then produces 3D-consistent images, and feature interpolation creates colors and shapes never seen in training. The paper reports that the output modification is essential because training does not converge without it, and that both the label embedding and the VGG discriminator improve FID scores. In quantitative comparisons, the method is said to perform on par with CLIP-NeRF and to exceed GRAF and GIRAFFE in PSNR and SSIM, while keeping model storage and training time constant as the number of scenes increases.

Load-bearing premise

The argument rests on the assumption that multiplying the shape and appearance latent codes by learned label embeddings gives one shared MLP enough capacity and separation to keep many distinct scenes from interfering with one another; the paper itself reports that FID worsens as classes and styles grow, so this capacity assumption is the fragile load-bearing premise.

Editorial extensions

If this is right

  • Storage and training cost for a generative 3D model no longer scale with the number of scenes; the same single MLP is reused for all classes and styles.
  • A user can explicitly choose object class and color at inference time, replacing the unpredictable latent-code sampling of GRAF.
  • Novel views can be rendered by changing camera pose, radius, or translation without retraining.
  • Linear interpolation between density or color array entries synthesizes shapes and colors absent from the training set.
  • The reported quality drop with more classes and styles marks a capacity limit of the shared MLP, so the approach is most effective when the number of scenes is moderate.

Reading between the lines

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

  • The same label-multiplication mechanism could plausibly extend to other attributes the paper lists as future features, such as texture, material, and illumination, by adding more embedding vectors and output slots, at the cost of the same entanglement-induced degradation.
  • Because the output modification is what makes training converge, a useful stress test would be to compare multiplicative label embeddings against additive conditioning or feature-wise modulation within this architecture; the paper does not run that comparison.
  • The constant-storage advantage over CLIP-NeRF becomes more attractive as the number of scenes grows, but the declining image quality suggests an upper bound on scene count before per-scene models become preferable; that crossover point is not measured in the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper introduces CtrlNeRF, a GRAF-based generative neural radiance field that aims to use a single MLP with shared weights to represent multiple scenes, conditioned on class labels (shape) and color labels (appearance). The generator is described as taking label-embedded latent codes and outputting density and color arrays indexed by labels, with a VGG-based discriminator as an auxiliary classifier. Experiments on a synthetic CARs dataset, NeRF-Synthetic, and LLFF report FID/KID, PSNR/SSIM, qualitative results for label-controlled generation, novel view synthesis, and interpolation, together with an ablation study.

Significance. If the central claim held, representing multiple scenes in one MLP with explicit control would be a useful efficiency result for 3D-aware generative models. The paper has strengths: it reports quantitative FID/KID per class and style, includes an ablation, and honestly documents degradation with scene count in Fig. 10, which is a falsifiable observation. However, the evidence does not currently establish the central claim: the scaling results show substantial quality loss, no per-scene baseline at matched storage is provided, and the PSNR/SSIM comparison to CLIP-NeRF is not a matched comparison. The contribution is incremental over GRAF, but the direction is reasonable.

major comments (4)
  1. [Table 2, §5.5] The PSNR/SSIM comparison in Table 2 is not a matched comparison: CLIP-NeRF is a per-scene reconstruction and manipulation method trained separately on posed images of each scene, whereas CtrlNeRF is a generative model with a single MLP across scenes. Reporting that CLIP-NeRF achieves higher PSNR/SSIM and then concluding that CtrlNeRF is 'on par' or 'marginally less impressive' is misleading because the methods solve different tasks. To support the multi-scene claim, the authors should compare against per-scene GRAF models (or a per-scene variant of CtrlNeRF) at matched total storage and report FID per scene, and use PSNR/SSIM only against equally generative baselines.
  2. [§5.1, Fig. 10, §6] The scaling experiments show FID rising steeply as the number of classes N and styles M increases; for example, the mean FID on LLFF is roughly 80 versus about 46.6 on CARs. The text in §5.5 asserts that the method stores multiple scenes 'without significantly sacrificing image quality,' but no comparison against per-scene baselines at matched storage is given, and §6 concedes that image quality diminishes as scene quantity and complexity increase. The central claim of efficient multi-scene representation therefore lacks quantitative support. Provide a matched-storage comparison (e.g., N per-scene GRAF models versus one CtrlNeRF) reporting FID and per-scene quality, and report error bars across seeds.
  3. [§5.4, Table 1] The ablation shows that removing the per-label output arrays (Model II) causes training not to converge, meaning these arrays are load-bearing. This indicates that the model is a shared backbone with per-class density heads and per-style color heads, rather than a fully shared representation. The contribution of the label-embedding multiplication (the input modification) is not isolated: there is no variant that keeps the output arrays but removes the multiplication, for example by using concatenation or no embedding at all. Please add such an ablation and clarify exactly which parameters are shared across scenes.
  4. [§3, Discriminator] The sentence 'To further improve image quality, we adopted posed images for training and replaced adversarial loss with reconstruction loss' contradicts the rest of the method: Equation 6 includes L_adv, Algorithm 1 updates both D_phi and D_vgg, and the positioning of CtrlNeRF as a GRAF-derived generative model suggests training without pose supervision. If posed images are indeed used, this is a major deviation that must be stated in the setup; if not, the sentence should be removed. Please clarify the exact training objective and data requirements.
minor comments (7)
  1. [§3, Output] M and N are swapped in the array definitions: the density array is written as [σ(i)]_{i=0}^{N-1} while M denotes the number of classes and N the number of styles, but labels i and j index classes and styles, respectively, in Algorithm 1 and Fig. 4; make the notation consistent.
  2. [Eq. (1), §4.3] Equation 1 has mismatched parentheses, and the FID score is attributed to 'Kanazawa et al.' although the standard reference is Heusel et al.; please correct both.
  3. [§3, References] Unresolved reference placeholders appear for pixelNeRF and VGG16 in Section 3, and 'GRIFFEE' should be 'GIRAFFE' throughout the text and figure captions.
  4. [§6] The statement that the authors 'successfully entangled and disentangled the label codes into and out of the latent code' is contradictory and should be rephrased to describe the intended conditioning mechanism.
  5. [§3, GRAF review] The GRAF description swaps the roles of z_s and z_a, calling z_a the shape code and z_s the appearance code, which conflicts with the notation in the abstract and Figure 2; please unify the notation.
  6. [Table 1, Figs. 8–10] FID and KID values are reported without standard deviations or the number of seeds; for a GAN evaluation, results from at least three training runs should be reported as mean ± std.
  7. [§5.5, Table 2] The paper does not specify the resolution, training iterations, or data used to train GRAF and GIRAFFE for the Table 2 comparison, so it is unclear whether the comparison is on the same scenes and settings; please document the protocol.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all reported outputs are forward renderings or explicitly defined interpolations from the learned radiance field, not fitted values renamed as predictions.

full rationale

CtrlNeRF is an empirical generative system rather than a derivation chain. Label-conditioned images are produced by feeding label-embedded latent codes through the MLP and volume-rendering the resulting density/color arrays (Eqs. 4 and 6), so class/style control is architectural conditioning, not a quantity fitted to a subset and then 'predicted' on a closely related subset. Novel-view results (Figs. 11-13) are direct forward passes at new camera poses through the same radiance field; no held-out pose is fit and then read off. The interpolation demos (Figs. 14-15) are explicitly defined as c=(1-λ)c[i]+λc[j], so the 'new' color/shape is a linear combination of already-rendered outputs; this is a construction rather than an independent prediction, which limits significance but is not circular. The ablation (Table 1) openly shows the per-label output arrays are necessary (Model II does not converge), a caveat on the 'single shared MLP' claim rather than a circular justification. The only self-reference ([48], by one author) is not cited as evidence for any premise. The stated limitation in Sec. 6 that image quality diminishes with scene quantity and complexity is an empirical scaling observation and further confirms the results are not forced by definition.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method relies on standard NeRF volume rendering, the GRAF framework, and two ad hoc assumptions: multiplication-based label conditioning and shared-MLP capacity. These are not derived or formally verified.

free parameters (6)
  • lambda1 = 2.0
    Weight for class loss in Equation 6, chosen by hand.
  • lambda2 = 3.0
    Weight for style loss in Equation 6, chosen by hand.
  • latent_code_dim = 128
    Dimensionality of shape and appearance codes, a design choice.
  • num_rays_per_image = 1024
    Number of rays sampled per image in experiments, fixed setting.
  • num_points_per_ray = 64
    Number of sample points per ray, fixed setting.
  • learning_rates = 0.0001 (D), 0.0005 (G)
    Optimizer learning rates for discriminator and generator.
assumptions (4)
  • standard math Volume rendering integral and stratified sampling are valid approximations of the radiance field equation.
    Used throughout Section 3 via Equations 3 and 4, inherited from NeRF.
  • domain assumption GRAF's generative radiance field pipeline is a sound starting point.
    The paper builds directly on GRAF's generator and patch discriminator.
  • ad hoc to paper Label-embedded latent codes obtained by multiplication provide sufficient conditioning for class and style control.
    Proposed in Section 3, no analysis shows why multiplication is better than concatenation or other conditioning.
  • ad hoc to paper A single shared-parameter MLP can represent multiple scenes without unacceptable interference.
    Core assumption of the method; the paper's own results show quality degradation with more scenes, so this is fragile.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CtrlNeRF: The Generative Neural Radiation Fields for the Controllable Synthesis of High-fidelity 3D-Aware Images." pith.science (2026). https://pith.science/paper/RLRDVPOK

@misc{pith2026241200754,
  author       = {Pith},
  title        = {Pith review of: CtrlNeRF: The Generative Neural Radiation Fields for the Controllable Synthesis of High-fidelity 3D-Aware Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RLRDVPOK}},
  note         = {Machine review of arXiv:2412.00754}
}
read the original abstract

The neural radiance field (NERF) advocates learning the continuous representation of 3D geometry through a multilayer perceptron (MLP). By integrating this into a generative model, the generative neural radiance field (GRAF) is capable of producing images from random noise z without 3D supervision. In practice, the shape and appearance are modeled by z_s and z_a, respectively, to manipulate them separately during inference. However, it is challenging to represent multiple scenes using a solitary MLP and precisely control the generation of 3D geometry in terms of shape and appearance. In this paper, we introduce a controllable generative model (i.e. \textbf{CtrlNeRF}) that uses a single MLP network to represent multiple scenes with shared weights. Consequently, we manipulated the shape and appearance codes to realize the controllable generation of high-fidelity images with 3D consistency. Moreover, the model enables the synthesis of novel views that do not exist in the training sets via camera pose alteration and feature interpolation. Extensive experiments were conducted to demonstrate its superiority in 3D-aware image generation compared to its counterparts.

Figures

Figures reproduced from arXiv: 2412.00754 by the authors.

Figure 1
Figure 1. The architecture of generative adversarial networks (GANs). G refers to a generator, and D refers to a binary discriminator. been successful in applications of reverse rendering, novel view synthesis, 3D object editing, digital human bodies, and image/video processing. The primary limitations of NERFs are that they require posed images for training and are unable to learn multiple scenes using a single MLP. By integ… view at source ↗
Figure 2
Figure 2. The framework of the generative neural radiation field (CtrlNeRF), which includes three main components: embedding, generator, and discriminator. control of the camera pose for novel view synthesis, and does not require posed images for training. However, a bottleneck is that one MLP merely represents a scene, resulting in high memory overhead in the case of multiple scenes. Fur￾thermore, GRAF cannot provide sophist… view at source ↗
Figure 4
Figure 4. The architecture of a conditional radiance field (MLP) comprises inputs of 𝑧 ′ 𝑠 and 𝑧 ′ 𝑎 , as well as 𝛾(𝑥) and 𝛾(𝑑). The output of the model consists of a volume density array 𝜎[] and color array c[]. label embedding technique is illustrated in Fig.4. Conditional Radiance Field: The inference for the volume density and color resembles that of the GRAF prototype. However, the MLP in the generator is conditional on … view at source ↗
Figures from the paper (10 more)
Figure 3
Figure 3. Figure 3: Scheme for incorporating label codes into a latent code through multiplication. Input: The latent code 𝑧 is typically separated into two components: shape code 𝑧𝑠 and appearance code 𝑧𝑎 . The text labels were initially translated into numerical labels (i=0, 1, 2, ... )…
Figure 5
Figure 5. Figure 5: Samples of the synthesized images (400x400) on CARs(I) dataset. i=0 i=1 i=2 i=3 Chair Fern Hot Dog Lego [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Samples of the synthesized images (400x400) on Synthetic(II) dataset. i=0 i=1 i=2 i=3 i=4 j=0 j=1 j=2 j=3 Real Hotjet Gray fern flower horns orchids room trex i=5 Inverse [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Samples of the synthesized images (504x378) on LLFF(III) dataset. model, it is evident that the image quality decreases as the number of scenes increases. The mean FID scores of the model trained on Datasets (I), (II), and (III) are presented in Fig.9. As shown in [PI…
Figure 8
Figure 8. Figure 8: The diagram of FID scores of the generated images grouped by class (i=0,1,2,3), and color (j=0,1,2,3) Dataset I Dataset II Dataset III 50 55 60 65 70 75 80 85 FID scores Mean FID Scores on three datasets [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: The mean FID scores of the generated images on CARs(I), Synthetic(II) and LLFF(III) datasets, respectively 1 2 3 4 5 Number of classes(N) 0 20 40 60 80 100 FID scores 1 2 3 4 5 Number of styles(M) 0 20 40 60 80 100 FID scores [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 13
Figure 13. Figure 13: The horizontal translation of the synthetic object with the distance d from -1.0 to 1.0. Here, d is the horizontal shift of the camera. 5.3. New Feature Synthesis via Linear Interpolation. As shown in Fig.14, the new color of the car, which is unseen in the training s…
Figure 14
Figure 14. Figure 14: The color of the car is synthesized via color linear interpolation. 𝜆 is a linear coefficient that varies from 0 to 1. 𝜆=0 𝜆=0.5 𝜆=1.0 [PITH_FULL_IMAGE:figures/full_fig_p009_14.png]
Figure 15
Figure 15. Figure 15: The shape of the car is altered via density linear interpolation. 𝜆 is a linear coefficient ranging from 0 to 1. also simulate other features, such as texture, material, and environmental illumination. As shown in Fig.15, the shape of the car can also be altered step …
Figure 16
Figure 16. Figure 16: Some synthesized images by GRAF, GIRAFFE, CLIP-NERF and our model on the Synthetic(II) dataset. [3] Mildenhall, Ben and Srinivasan, Pratul P. and Tancik, Matthew and Barron, Jonathan T. and Ramamoorthi, Ravi and Ng, Ren. NeRF: Representing Scenes as Neural Radiance Fi…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 23 canonical work pages

  1. [1]

    DOI:10.23919/TST.2017.8195348

    Wu,XianandXu,KunandHall,Peter.Asurveyofimagesynthesisand editing with generative adversarial networks Tsinghua Science Tech- nology, 22(006),660-674, 2018. DOI:10.23919/TST.2017.8195348

    work page arXiv 2018

  2. [2]

    GRAF: Generative Radiance Fields for 3D-Aware Image Synthesis

    Schwarz, Katja and Liao, Yiyi and Niemeyer, Michael and Geiger, Andreas. GRAF: Generative Radiance Fields for 3D-Aware Image Synthesis, arXiv e-prints, 2020. DOI:10.48550/arXiv.2007.02442. : Page 9 of 12 Table 1 The results of FID/KID scores in the ablation studies. Model/Class Classic Sporty Roadster Wagon Mean FID/KID Model I 68.83 71 .07 80 .18 74 .92 ...

    work page Pith review arXiv doi:10.48550/arxiv.2007.02442 2020

  3. [3]

    and Tancik, Matthew and Barron, Jonathan T

    Mildenhall, Ben and Srinivasan, Pratul P. and Tancik, Matthew and Barron, Jonathan T. and Ramamoorthi, Ravi and Ng, Ren. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis, arXiv e-prints, 2020. DOI:10.48550/arXiv.2003.08934

    arXiv doi:10.48550/arxiv.2003.08934 2020

  4. [4]

    Goodfellow,IanJ.andPouget-Abadie,JeanandMirza,MehdiandXu, BingandWarde-Farley,DavidandOzair,SherjilandCourville,Aaron andBengio,Yoshua.GenerativeAdversarialNetworks,.arXive-prints,

    work page

  5. [5]

    Recent Progress on Generative Adver- sarial Networks (GANs): A Survey, IEEE Access, 36322-36333,2019

    Pan,ZhaoqingandYu,WeijieandYi,XiaokaiandKhan,Asifullahand Yuan, Feng and Zheng, Yuhui. Recent Progress on Generative Adver- sarial Networks (GANs): A Survey, IEEE Access, 36322-36333,2019. DOI:10.1109/ACCESS.2019.2905015

    work page arXiv 2019

  6. [6]

    UnsupervisedRepresentationLearningwithDeepConvolutionalNeu- ral Network for Remote Sensing Images, International Conference on Image & Graphics, 2017

    Yu, Yang and Gong, Zhiqiang and Zhong, Ping and Shan, Jiaxin. UnsupervisedRepresentationLearningwithDeepConvolutionalNeu- ral Network for Remote Sensing Images, International Conference on Image & Graphics, 2017

    work page 2017

  7. [7]

    Conditional Generative Adversarial Nets, Computer Science, 2672-2680, 2014

    Mirza, Mehdi and Osindero, Simon. Conditional Generative Adversarial Nets, Computer Science, 2672-2680, 2014. DOI:10.48550/arXiv.1411.1784

    arXiv doi:10.48550/arxiv.1411.1784 2014

  8. [8]

    Condi- tionalImageSynthesisWithAuxiliaryClassifierGANs,arXive-prints,

    Odena, Augustus and Olah, Christopher and Shlens, Jonathon. Condi- tionalImageSynthesisWithAuxiliaryClassifierGANs,arXive-prints,

    work page

Show all 55 references
  1. [9]

    InfoGAN:InterpretableRepre- sentationLearningbyInformationMaximizingGenerativeAdversarial Nets, arXiv e-prints,2016

    Chen, Xi and Duan, Yan and Houthooft, Rein and Schulman, John andSutskever,IlyaandAbbeel,Pieter. InfoGAN:InterpretableRepre- sentationLearningbyInformationMaximizingGenerativeAdversarial Nets, arXiv e-prints,2016. DOI:10.48550/arXiv.1606.03657

    2016 doi arXiv

  2. [10]

    Wasserstein generativeadversarialnetworks,InternationalConferenceonMachine Learning, 2017

    Arjovsky,MartinandChintala,SoumithandBottou,On. Wasserstein generativeadversarialnetworks,InternationalConferenceonMachine Learning, 2017. DOI:10.1088/1742-6596/2586/1/012157

    2017 doi

  3. [11]

    Improved Training of Wasser- stein GANs, 2017

    Gulrajani, Ishaan and Ahmed, Faruk and Arjovsky, Martin and Du- moulin, Vincent and Courville, Aaron. Improved Training of Wasser- stein GANs, 2017. DOI:10.48550/arXiv.1704.00028. : Page 10 of 12 Table 2 The quantitative assessment of our model with state-of-the-art methods in...

    2017 doi arXiv

  4. [12]

    On the regularization of Wasserstein GANs, 2017

    Petzka, Henning and Fischer, Asja and Lukovnicov, Denis. On the regularization of Wasserstein GANs, 2017. DOI:10.48550/arXiv.1709.08894

    2017 doi arXiv

  5. [13]

    Progressive Growing of GANs for Improved Quality, Stability, and Variation, 2017

    Karras,TeroandAila,TimoandLaine,SamuliandLehtinen,Jaakko. Progressive Growing of GANs for Improved Quality, Stability, and Variation, 2017. DOI:10.48550/arXiv.1710.10196

    2017 doi arXiv

  6. [14]

    IEEE ICCV, 2017

    Zhang, Han and Xu, Tao and Li, Hongsheng and Zhang, Shaot- ing and Wang, Xiaogang and Huang, Xiaolei and Metaxas, Dim- itris. IEEE ICCV, 2017. StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks, DOI:10.1109/ICCV.2017.629

    2017 doi

  7. [15]

    IEEE Transactions on Pattern Analysis and Ma- chine Intelligence,PP,99, 2017

    StackGAN++: Realistic Image Synthesis with Stacked Generative Adversarial Networks, Han, Zhang and Tao, Xu and Hongsheng, Li and Shaoting, Zhang and Xiaogang, Wang and Xiaolei, Huang and Metaxas, Dimitris N. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence,PP,9...

    2017

  8. [16]

    High-ResolutionImage Synthesis and Semantic Manipulation with Conditional GANs, 2017

    Wang, Ting Chun and Liu, Ming Yu and Zhu, Jun Yan and Tao, AndrewandKautz,JanandCatanzaro,Bryan. High-ResolutionImage Synthesis and Semantic Manipulation with Conditional GANs, 2017. DOI:10.48550/arXiv.1711.11585

    2017 doi arXiv

  9. [17]

    A Style-Based Gen- erator Architecture for Generative Adversarial Networks, Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, 2019

    Karras, Tero and Laine, Samuli and Aila, Timo. A Style-Based Gen- erator Architecture for Generative Adversarial Networks, Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, 2019. DOI:10.1109/TPAMI.2020.2970919

    2019

  10. [18]

    High-Fidelity Synthesis with Disentangled Representation,

    Lee,WonkwangandKim,DonggyunandHong,SeunghoonandLee, Honglak. High-Fidelity Synthesis with Disentangled Representation,

  11. [19]

    DOI:10.48550/arXiv.2003.03461

    Nie,WeiliandKarras,TeroandGarg,AnimeshandDebhath,Shoub- hik and Anandkumar, Anima, Semi-Supervised StyleGAN for Disen- tanglement Learning, 2020. DOI:10.48550/arXiv.2003.03461

    2020 doi arXiv

  12. [20]

    and Sohn, K

    Reed, S. and Sohn, K. and Zhang, Y. and Lee, H. Learning to Dis- entangle Factors of Variation with Manifold Interaction, International Conference on Machine Learning, 2014

    2014

  13. [21]

    Advances in Neural Rendering, arXiv e-prints,

    Tewari, Ayush and Thies, Justus and Mildenhall, Ben and Srini- vasan, Pratul and Tretschk, Edgar and Wang, Yifan and Lassner, Christoph and Sitzmann, Vincent and Martin-Brualla, Ricardo and Lombardi, Stephen. Advances in Neural Rendering, arXiv e-prints,

  14. [22]

    M and Weston, Nick

    Brock, Andrew and Lim, Theodore and Ritchie, J. M and Weston, Nick. Generative and Discriminative Voxel Modeling with Convolu- tional Neural Networks, Computer Science, 2016

    2016

  15. [23]

    Rezende, Danilo Jimenez and Eslami, S. M. Ali and Mohamed, Shakir and Battaglia, Peter and Jaderberg, Max and Heess, Nicolas. Unsupervised Learning of 3D Structure from Images In Advances in Neural Information Processing Systems (NeurIPS), 2016

    2016

  16. [24]

    OctNet: Learning Deep 3D Representations at High Resolutions, 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR),

    Riegler,GernotandUlusoy,AliOsmanandGeiger,Andreas. OctNet: Learning Deep 3D Representations at High Resolutions, 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR),

    2017

  17. [25]

    andTenenbaum, Joshua B

    Wu, Jiajun and Zhang, Chengkai and Xue, Tianfan and Freeman, William T. andTenenbaum, Joshua B. Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling, In AdvancesinNeuralInformationProcessingSystems(NeurIPS),2016. DOI:10.48550/arXiv.1610.07584

    2016 doi arXiv

  18. [26]

    AtlasNet: A Papier-Mché Approach to Learning 3D Surface Generation, In Proc

    Groueix, Thibault and Fisher, Matthew and Kim, Vladimir G and Russell, Bryan C and Aubry, Mathieu. AtlasNet: A Papier-Mché Approach to Learning 3D Surface Generation, In Proc. IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR), 2018. DOI:10.48550/arXiv.1802.05384

    2018 doi arXiv

  19. [27]

    Deep March- ing Cubes: Learning Explicit Surface Representations, IEEE/CVF Conference on Computer Vision and Pattern Recognition,2018

    Liao, Yiyi and Simon Donné and Geiger, Andreas. Deep March- ing Cubes: Learning Explicit Surface Representations, IEEE/CVF Conference on Computer Vision and Pattern Recognition,2018. DOI:10.1109/CVPR.2018.00308

    2018

  20. [28]

    Deep Mesh Reconstruction From Single RGB Images via Topology Modification Networks, IEEE/CVF International Conference on Computer Vision (ICCV), 2020

    Pan, Junyi and Han, Xiaoguang and Chen, Weikai and Tang, Jiapeng and Jia, Kui. Deep Mesh Reconstruction From Single RGB Images via Topology Modification Networks, IEEE/CVF International Conference on Computer Vision (ICCV), 2020. DOI:10.1109/ICCV.2019.01006

    2020

  21. [29]

    Pixel2Mesh: Generating 3D Mesh Models from Single RGB Images, In Proc

    Wang, Nanyang and Zhang, Yinda and Li, Zhuwen and Fu, Yanwei and Liu, Wei and Jiang, Yu Gang. Pixel2Mesh: Generating 3D Mesh Models from Single RGB Images, In Proc. of the European Conf. on ComputerVision(ECCV),2018. DOI:10.1007/978-3-030-01252-6_4

    2018 doi

  22. [31]

    title=Convolutional Occupancy Networks, Peng, Songyou and Niemeyer, Michael and Mescheder, Lars and Pollefeys, Marc and Geiger, Andreas. In Proc. of the European Conf. on Computer Vision (ECCV),2020. DOI:10.1007/978-3-030-58580-8_31

    2020 doi

  23. [33]

    Differentiable Volumetric Rendering: Learning Implicit3DRepresentationsWithout3DSupervision,IEEE/CVFCon- ference on Computer Vision and Pattern Recognition (CVPR), 2020

    Niemeyer, Michael and Mescheder, Lars and Oechsle, Michael and Geiger, Andreas. Differentiable Volumetric Rendering: Learning Implicit3DRepresentationsWithout3DSupervision,IEEE/CVFCon- ference on Computer Vision and Pattern Recognition (CVPR), 2020. DOI:10.1109/CVPR42600.2020.00356

    2020

  24. [34]

    SceneRepresentationNetworks:Continuous3D-Structure-Aware NeuralSceneRepresentations,InAdvancesinNeuralInformationPro- cessing Systems (NeurIPS), 2019

    Sitzmann, Vincent and Zollhfer, Michael and Wetzstein, Gor- don. SceneRepresentationNetworks:Continuous3D-Structure-Aware NeuralSceneRepresentations,InAdvancesinNeuralInformationPro- cessing Systems (NeurIPS), 2019. DOI:10.48550/arXiv.1906.01618

    2019 doi arXiv

  25. [35]

    Pix2Vox: Context-Aware 3D Reconstruction From Single and Multi-View Images, Inter- national Conference on Computer Vision (ICCV), IEEE, 2020

    Xie, Haozhe and Yao, Hongxun and Sun, Xiaoshuai and Zhou, Shangchen and Zhang, Shengping. Pix2Vox: Context-Aware 3D Reconstruction From Single and Multi-View Images, Inter- national Conference on Computer Vision (ICCV), IEEE, 2020. DOI:10.1109/ICCV.2019.00278

    2020

  26. [36]

    Es- caping Plato’s Cave: 3D Shape From Adversarial Rendering, In- ternational Conference on Computer Vision (ICCV), IEEE, 2019

    Henzler, Philipp and Mitra, Niloy and Ritschel, Tobias. Es- caping Plato’s Cave: 3D Shape From Adversarial Rendering, In- ternational Conference on Computer Vision (ICCV), IEEE, 2019. DOI:10.1109/ICCV.2019.01008

    2019

  27. [37]

    HoloGAN: Unsupervised learning of 3D representations from natural images, ICCV, 2019

    Nguyen-Phuoc, Thu and Li, Chuan and Theis, Lucas and Richardt, Christian and Yang, Yong Liang. HoloGAN: Unsupervised learning of 3D representations from natural images, ICCV, 2019. DOI:10.1109/ICCVW.2019.00255

    2019

  28. [38]

    BlockGAN: Learning 3D Object-aware Scene Representations from Unlabelled Images, 2020

    Nguyen-Phuoc, Thu and Richardt, Christian and Mai, Long and Yang, Yong Liang and Mitra, Niloy. BlockGAN: Learning 3D Object-aware Scene Representations from Unlabelled Images, 2020. DOI:10.48550/arXiv.2002.08988

    2020 doi arXiv

  29. [40]

    GIRAFFE: Representing ScenesasCompositionalGenerativeNeuralFeatureFields,IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

    Niemeyer, Michael and Geiger, Andreas. GIRAFFE: Representing ScenesasCompositionalGenerativeNeuralFeatureFields,IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

  30. [41]

    and Arbel, Michael and Gretton, Arthur

    Bińkowski, Mikoaj and Sutherland, Dougal J. and Arbel, Michael and Gretton, Arthur. Demystifying MMD GANs, 2018. DOI:10.48550/arXiv.1801.01401

    2018 doi arXiv

  31. [42]

    and Kowalski, Marek and Johnson, Matthew and Shotton, Jamie and Valentin, Julien

    Garbin, Stephan J. and Kowalski, Marek and Johnson, Matthew and Shotton, Jamie and Valentin, Julien. FastNeRF: High-Fidelity Neural Rendering at 200FPS, 2021. DOI:10.48550/arXiv.2103.10380

    2021 doi arXiv

  32. [43]

    NeuralSceneFlowFieldsforSpace-TimeViewSynthesisofDynamic Scenes, 2020

    Li,ZhengqiandNiklaus,SimonandSnavely,NoahandWang,Oliver. NeuralSceneFlowFieldsforSpace-TimeViewSynthesisofDynamic Scenes, 2020. DOI:10.48550/arXiv.2011.13084

    2020 doi arXiv

  33. [44]

    pixelNeRF: Neural Radiance Fields from One or Few Images, 2020

    Yu,AlexandYe,VickieandTancik,MatthewandKanazawa,Angjoo. pixelNeRF: Neural Radiance Fields from One or Few Images, 2020. DOI:10.48550/arXiv.2012.02190

    2020 doi arXiv

  34. [45]

    and Martin-Brualla, Ricardo and Snavely, Noah and Funkhouser, Thomas

    Wang, Qianqian and Wang, Zhicheng and Genova, Kyle and Srinivasan, Pratul and Zhou, Howard and Barron, Jonathan T. and Martin-Brualla, Ricardo and Snavely, Noah and Funkhouser, Thomas. IBRNet: Learning Multi-View Image-Based Rendering,

  35. [47]

    StylizedNeRF: Consistent 3D Scene Stylization as Stylized NeRF via 2D-3D Mutual Learning, IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai and Lin Gao. StylizedNeRF: Consistent 3D Scene Stylization as Stylized NeRF via 2D-3D Mutual Learning, IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  36. [48]

    Liu and T

    J. Liu and T. H. Lin. A framework for the synthesis of X-ray security inspectionimagesbasedongenerativeadversarialnetworks,IEEEAc- cess, 11, 63751-63760, 2023. DOI:10.1109/ACCESS.2023.3288087

    2023

  37. [49]

    Image quality assessment: From error visibility to structural similarity, 2013

    Wang,Z.andPan,W.andCuppens-Boulahia,N.andCuppens,F.and Roux, C. Image quality assessment: From error visibility to structural similarity, 2013

    2013

  38. [50]

    HeadNeRF: A Real-time NeRF-based Parametric Head Model, IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    Hong, Yang and Peng, Bo and Xiao, Haiyao and Liu, Ligang and Zhang, Juyong. HeadNeRF: A Real-time NeRF-based Parametric Head Model, IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. DOI:10.48550/arXiv.2112.05637

    2022 doi arXiv

  39. [51]

    GRAM: Generative Radiance Manifolds for 3D-Aware Image Gener- ation, IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR), 2022

    Deng, Yu and Yang, Jiaolong and Xiang, Jianfeng and Tong, Xin. GRAM: Generative Radiance Manifolds for 3D-Aware Image Gener- ation, IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR), 2022. DOI:10.48550/arXiv.2112.08867

    2022 doi arXiv

  40. [52]

    DOI:10.1109/CVPR46437.2021.00466

    2021

  41. [53]

    StyleNeRF: A Style-based 3D-Aware Generator for High-resolution Image Synthesis, 2021

    Gu, Jiatao and Liu, Lingjie and Wang, Peng and Theobalt, Christian. StyleNeRF: A Style-based 3D-Aware Generator for High-resolution Image Synthesis, 2021. DOI:10.48550/arXiv.2110.08985

    2021 doi arXiv

  42. [59]

    CLIP-NeRF: Text-and-Image Driven Manipulation of Neural Radiance Fields, arXiv e-prints, 2021

    Wang, Can and Chai, Menglei and He, Mingming and Chen, Dongdong and Liao, Jing. CLIP-NeRF: Text-and-Image Driven Manipulation of Neural Radiance Fields, arXiv e-prints, 2021. DOI:10.48550/arXiv.2112.05139. : Page 12 of 12

    2021 doi arXiv

  43. [2014]

    https://arxiv.org/pdf/1406.2661

    arXiv

  44. [2016]

    DOI:10.48550/arXiv.1610.09585

    doi arXiv

  45. [2017]

    DOI:10.1109/CVPR.2017.701

    2017 doi

  46. [2020]

    DOI:10.1007/978-3-030-58574-7_10

    doi

  47. [2021]

    DOI:10.48550/arXiv.2111.05849

    doi arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.