Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Fractal Flow: Hierarchical and Interpretable Normalizing Flow via Topic Modeling and Recursive Strategy

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A normalizing-flow architecture with a topic-modeled latent space and recursively nested coupling layers aims to improve both density estimation and interpretability.

desk verdict An interesting but under-specified architecture: the density math is sound and the topic-model latent idea has promise, but the empirical gain is confounded by the CNN subnetwork and the fractal layer is never actually specified. read the letter →

arxiv 2508.19750 v1 pith:LA5XRLSF submitted 2025-08-27 stat.ML cs.LG

Binhui Zhang , Jianwei Ma This is my paper
classification stat.MLcs.LG
keywords normalizingflowsdensityestimationlatentDirichletallocationKolmogorov-Arnoldnetworksfractalgenerativemodelsinterpretablespacecontrollablegenerationaffinecouplinglayers
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

Fractal Flow is a normalizing-flow architecture that tries to give a flow model a structured latent space without sacrificing density accuracy. The central claim is that two modifications—a Kolmogorov–Arnold-inspired latent prior whose mixture weights are organized by Latent Dirichlet Allocation, and a recursive 'fractal' arrangement of affine coupling layers operating at progressively smaller spatial scales—make the flow more interpretable and more accurate. In the paper's experiments the convolutional Fractal Flow reaches 1.20 bits per dimension on MNIST and 2.45 on FashionMNIST, against 1.77 and 3.04 for the authors' RealNVP baseline, while unsupervised latent topics separate shoes from clothing. The paper frames this as an extension of RealNVP-style coupling flows, not a replacement of the flow paradigm.

What carries the argument

The load-bearing object is the 'fractal coupling layer': an affine coupling layer whose masked blocks are themselves re-modeled by smaller-scale coupling layers at decreasing spatial resolutions, terminating in a KAN-inspired latent layer that models each dimension with grid-based Gaussian components and Dirichlet-distributed weights. The LDA-structured mixture is the mechanism for interpretability and controllable generation; the recursive decomposition is the mechanism for expressiveness and parameter efficiency. The architecture therefore combines two inductive biases: a topic-structured prior over latent codes and a self-similar, coarse-to-fine transformation path.

What would settle it

Run the convolutional Fractal Flow with the fractal recursion flattened into an equal-capacity, non-recursive stack of affine coupling layers while keeping the LDA prior and parameter budget fixed; if bits per dimension stays near 1.20 on MNIST and 2.45 on FashionMNIST, the recursive coupling is not responsible for the reported gain.

Watch

Extended reading notes

Core claim

The paper argues that the latent space of a normalizing flow can itself be the object of probabilistic modeling. Its structured prior writes the data density as a finite mixture of Gaussians in the latent space, with each outer component playing the role of a semantic topic; both the topic weights and the per-dimension grid weights are drawn from Dirichlet distributions, which is the LDA ingredient. The mean and variance of each Gaussian are tied to grid-based functions inspired by Kolmogorov–Arnold Networks, so the mixture is parameter-efficient and each latent dimension can be read as a small Gaussian mixture. On top of that prior, the paper places fractal coupling layers: each intermediat

Load-bearing premise

The load-bearing premise is that after each transformation step the latent space can be cut into independent blocks and recursively modeled by smaller-scale flows at decreasing spatial resolutions, and that this yields a valid, more expressive normalizing flow; the paper states this decomposition verbally and in figures but does not derive it, and the reported gains are not ablated separately from the convolutional subnetworks.

Editorial extensions

If this is right

  • If the reported numbers hold, flow-based density estimation can have a semantically organized latent space without paying an accuracy penalty relative to the RealNVP-style baseline.
  • The LDA topic layer gives a control handle for generation: samples drawn from different covariance or mean topics land in different semantic clusters, enabling label-free conditional sampling.
  • The recursive coupling design introduces a coarse-to-fine bias that is a natural fit for data with self-similar structure, and the paper demonstrates this qualitatively on seismic signals.
  • Because the recursion is applied on top of base affine coupling layers, it can be combined with other coupling-layer improvements rather than competing with them.
  • The paper's stated future directions—optimal-transport-based coupling and combination with diffusion models—would extend the same structured-prior/fractal idea to other generation paradigms.

Reading between the lines

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

  • Holding the subnetwork family fixed and varying only whether the coupling layers are recursively nested would isolate the recursive contribution; the paper does not report that ablation, although it compares subnetwork types.
  • Because the LDA prior only constrains the latent mixture, the same topic-structured prior could be dropped into a VAE or latent diffusion model for label-free semantic control.
  • On self-similar data such as seismic traces, the coarse-to-fine recursion may serve as an implicit prior for compression or reconstruction, but that use would need quantitative benchmarks rather than the qualitative samples shown.
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 / 5 minor

Summary. The paper proposes Fractal Flow, a normalizing-flow architecture that combines a structured latent prior inspired by Kolmogorov-Arnold Networks (KANs) and Latent Dirichlet Allocation (LDA) with a recursive 'fractal coupling layer' design. The latent prior is a mixture of factorized Gaussians over flow-transformed variables, with Dirichlet-distributed mixture weights and fixed grid means. The recursive component is described as partitioning the latent space into independent blocks and applying smaller-scale flows at decreasing spatial resolutions. Experiments report bits/dim on MNIST and FashionMNIST, plus qualitative CIFAR-10 automobile and seismic results, claiming that a CNN-based variant reaches 1.20 and 2.45 bits/dim versus 1.77 and 3.04 for a custom RealNVP baseline. The paper also claims improved latent clustering, interpretability, and controllable generation.

Significance. The structured latent prior formulation in Eqs. (2)-(9) is a reasonable way to attach mixture/LDA priors to normalizing flows, and the closed-form expected density in Eq. (9) is a useful construction. If the architecture were fully specified and evaluated against standard benchmarks with proper ablations, the LDA-structured prior could be a worthwhile contribution. However, as presented, the distinguishing 'fractal coupling layer' is not formally defined, and the empirical evidence does not isolate its contribution. The authors honestly acknowledge partial interpretability failures in Section 3.2 ('the model fails to consistently capture certain other categories'), which tempers the interpretability claim, but this honesty does not compensate for the missing methodological details. In its current form the central claims of architectural novelty and superior estimation accuracy are not verifiable.

major comments (3)
  1. [§2.4, Eq. (10), Fig. 5] The proposed fractal coupling layer is never defined. Section 2.4 states that after each transformation step the latent space is partitioned into independent blocks, each modeled by a smaller-scale flow at decreasing spatial resolutions (32x32, 16x16, etc.), but no equations specify the block partition, per-block transformations, parameter sharing, recursion depth, or how the resolution changes are composed invertibly. The only concrete construction is the standard affine coupling layer in Eq. (10). This is load-bearing because Table 1 attributes the reported accuracy gains to 'incorporating the fractal coupling layers,' yet the model cannot be reproduced or formally checked.
  2. [Table 1, §3.1] The empirical evidence for 'superior estimation accuracy' is confounded. Only a single custom RealNVP baseline is compared, with no parameter counts, error bars, or standard published benchmarks. LDANF (MLP) is essentially tied with RealNVP (1.76 vs 1.77 on MNIST; 3.03 vs 3.04 on FashionMNIST) and LDAFNF (KAN) is worse (1.80/3.24). Only LDAFNF (CNN) improves markedly. Since the difference among LDAFNF variants is the subnetwork type, the gain could be due to convolutional subnetworks rather than the fractal recursion or LDA structure. An ablation that keeps the subnetwork fixed and toggles the fractal layers is necessary before the headline claim can be accepted.
  3. [§2.1, Eqs. (1)-(2)] The KAT connection is presented as an integration, but the derivation is purely interpretive: the text says 'we interpret' and 'we reformulate,' and Eq. (2) does not follow from Eq. (1) in any stated formal sense. KAT is a representation theorem for continuous functions, whereas Eq. (2) is a mixture density with a Jacobian determinant; no argument is given for why the outer sum of univariate functions should map to a Gaussian mixture over transformed variables. If the contribution is an analogy, the paper should say so; if it is a mathematical connection, a formal statement is needed. As written, KAT supplies motivation rather than a proven component of the method.
minor comments (5)
  1. [Eq. (10)] The affine coupling expression has mismatched parentheses: y_transformed = x_transformed ⊙ exp(s(x_masked) + t(x_masked) is missing a closing parenthesis and appears incomplete. The standard RealNVP form is x ⊙ exp(s(x_masked)) + t(x_masked).
  2. [Abstract] The abstract says experiments on CIFAR-10 and geophysical data, but the paper only reports the automobile class of CIFAR-10 and qualitative seismic results with no likelihood evaluation. The wording overstates the empirical scope.
  3. [§2.3, Eq. (9)] Equation (9) is called the 'expected log-likelihood' but it actually computes the expected density E_{π,w} p(x | π, w). Please clarify the training objective and how the closed-form expectation is used, including how Dirichlet samples are incorporated during generation.
  4. [Fig. 5] The caption in Figure 5 refers to 'Fractal Generative Model' rather than 'Fractal Flow.' Clarify the relationship to FGMs and avoid naming the architecture differently from the paper title.
  5. [Table 1] The column header contains a typo ('F ashionMNIST'). More importantly, the table does not report parameter counts or standard deviations, despite the claim of 'controlled parameter sizes' in Section 3.1.

Circularity Check

1 steps flagged · score 4.0 of 10

The KAT 'derivation' is a post-hoc relabeling of a standard Gaussian-mixture latent prior; no fitted-value or self-citation circularity is present, but the claimed first-principles KAT link reduces to renaming.

  1. renaming known result [Section 2.1, bullet list after Eq. (4)]
    "1. Σ_{k=1}^K w_k |det(∂z/∂x)| / ((2π)^{D/2}|Σ_k|^{1/2}) exp(·) corresponds to the outer summation Σ_{q=1}^{2n+1} Φ_q(·) in the KAT. 2. Σ_{d=1}^D −(z_d−μ_{k,d})^2 / (2σ^2_{k,d}) corresponds to the inner summation Σ_{p=1}^n ϕ_{q,p}(·) in the KAT."

    The KAT representation theorem (Eq. 1) is a statement about representing continuous functions by sums of univariate functions; it imposes no density form. Eq. (2)/(4) is just the standard change-of-variables density with a Gaussian-mixture latent prior, a construction that predates this paper (the paper itself cites Izmailov et al. and Yao et al. for GMM/hierarchical priors in flows). The two bullet correspondences map the mixture sum to the KAT outer sum and the per-dimension Gaussian exponent to the KAT inner sum. This is a term-by-term notational relabeling, not a derivation. Therefore the claim that the structured latent prior is 'reformulated from KAT' is a post-hoc renaming of an assumed Gaussian-mixture prior; the KAT adds no constraint and the 'first-principles' status of the prior

full rationale

The empirical evaluation is self-contained: NLL comparisons in Table 1 are against the authors' own RealNVP reimplementation and their own LDA/LDAFNF variants, rather than an external benchmark, so there is no fitted parameter renamed as a prediction and no self-citation chain that forces the main accuracy claim. The citations to FGMs (Li et al.) and GLF (Xiao et al.) are external and inspirational, not load-bearing self-citations; the seismic-data citation to Yu and Ma is a data/domain reference. The one genuine circularity-like defect is the KAT reformulation: Eq. (2)/(4) is a standard Gaussian-mixture latent prior in a normalizing flow, and the mapping of its terms onto the KAT outer/inner sums is a relabeling, making the 'KAT-based structured latent prior' an after-the-fact interpretation rather than a derivation. Separately—but not as constructional circularity—Section 2.4 never specifies the fractal coupling construction (no equations for the block partition, per-block transformations, parameter sharing, recursion depth, or how decreasing spatial resolutions are composed invertibly), and Table 1 confounds the fractal component with the CNN subnetwork: LDAFNF (MLP) improves less, and LDAFNF (KAN) is worse than the baseline. These are missing-support and attribution problems that lower confidence in the 'superior estimation accuracy' claim, but they are not themselves equivalence-by-construction loops. Overall score 4 reflects the central KAT derivation reducing to a renaming of a known GMM latent prior, while the rest of the architecture still has independent empirical content.

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

The central claim relies on several hand-chosen architectural hyperparameters (Q, M, the fixed grid means, and the number of fractal recursion levels), plus the assumption that a factored Gaussian mixture prior with Dirichlet-distributed weights is expressive enough for the target data. The fractal decomposition is the least specified premise: the paper gives no formal construction or validity proof.

free parameters (4)
  • Number of outer components Q
    Number of Gaussian components in the outer mixture, chosen by hand; not reported in the paper (Eq. 5, Section 2.2).
  • Number of inner components M
    Number of Gaussian components per dimension in the inner mixture; chosen by hand, not reported (Eq. 5).
  • Dirichlet hyperparameters alpha, beta
    Learnable concentration parameters for the topic-weight Dirichlet distributions (Eqs. 7-8); these are fitted during training and directly control the mixture weights used in the density estimate (Eq. 9).
  • Fixed grid means mu_{qdm}
    Means are stated to be fixed and non-learnable to preserve grid structure (Section 2.3), but their values are never specified; they are effectively chosen hyperparameters that define the entire latent grid.
assumptions (5)
  • standard math Kolmogorov-Arnold representation theorem holds for continuous functions (Eq. 1)
    Invoked in Section 2.1 as the inspiration for the structured prior; the paper does not use it as a quantitative guarantee.
  • domain assumption The invertible transformation z=T(x) is a valid normalizing flow and can be parameterized by neural networks with tractable Jacobians
    Standard NF assumption; the paper relies on affine coupling layers to satisfy it (Section 2.1).
  • domain assumption Latent dimensions are statistically independent after the transformation (Eq. 4)
    The paper introduces this independence assumption to make the mixture tractable; it is not proven and constrains the prior.
  • ad hoc to paper Fixing the grid means and treating weights as Dirichlet random variables yields an expressive enough latent prior
    The paper asserts this addresses mode collapse and improves diversity (Section 2.3), but provides no analysis or evidence beyond qualitative samples.
  • ad hoc to paper Recursive decomposition of the latent space into independent blocks with decreasing resolution preserves the validity of the flow and improves accuracy
    The fractal design is described in prose in Section 2.4; no theorem or precise construction is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fractal Flow: Hierarchical and Interpretable Normalizing Flow via Topic Modeling and Recursive Strategy." pith.science (2026). https://pith.science/paper/LA5XRLSF

@misc{pith2026250819750,
  author       = {Pith},
  title        = {Pith review of: Fractal Flow: Hierarchical and Interpretable Normalizing Flow via Topic Modeling and Recursive Strategy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LA5XRLSF}},
  note         = {Machine review of arXiv:2508.19750}
}
read the original abstract

Normalizing Flows provide a principled framework for high-dimensional density estimation and generative modeling by constructing invertible transformations with tractable Jacobian determinants. We propose Fractal Flow, a novel normalizing flow architecture that enhances both expressiveness and interpretability through two key innovations. First, we integrate Kolmogorov-Arnold Networks and incorporate Latent Dirichlet Allocation into normalizing flows to construct a structured, interpretable latent space and model hierarchical semantic clusters. Second, inspired by Fractal Generative Models, we introduce a recursive modular design into normalizing flows to improve transformation interpretability and estimation accuracy. Experiments on MNIST, FashionMNIST, CIFAR-10, and geophysical data demonstrate that the Fractal Flow achieves latent clustering, controllable generation, and superior estimation accuracy.

Figures

Figures reproduced from arXiv: 2508.19750 by the authors.

Figure 1
Figure 1. Visualization of a structured latent space using the Structured Latent Priors. The [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. From left to right: RealNVP-based latent space, RealNVP-based MOON’s prob [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Neural network implementation of the structured latent space in a KAT-based [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Statistical LDA-based modeling of the structured latent space. The latent space [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Network architecture of Fractal Generative Model based on affine coupling layers. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Ablation study of the LDA layer and fractal coupling layer, along with compar [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: The top row shows samples from MNIST and FashionMNIST. The bottom row [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: The top row shows random samples generated from two different covariance top [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Visualization of the output at each transformation step on MNIST. [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Visualization of the output at each transformation step on FashionMNIST. [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Random samples generated on the automobile class of CIFAR-10. [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Random samples generated on the seismic dataset. [PITH_FULL_IMAGE:figures/full_fig_p019_12.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Learning the Universe: Posterior Reliability of Neural Generative Models in High-Dimensional Field-Level Inference of Cosmic Initial Conditions

    astro-ph.CO 2026-06 unverdicted novelty 6.0 of 10

    Generative models for cosmological field-level inference can reproduce posterior means and cross-correlations yet fail to capture correct uncertainty geometry when validated against HMC reference samples.

Reference graph

Works this paper leans on

33 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [1]

    Fractal structures in nonlinear dynamics

    Jacobo Aguirre, Ricardo L Viana, and Miguel AF Sanju \'a n. Fractal structures in nonlinear dynamics. Reviews of Modern Physics, 81 0 (1): 0 333--386, 2009

    work page 2009

  2. [2]

    Neural flow diffusion models: Learnable forward process for improved diffusion modelling

    Grigory Bartosh, Dmitry P Vetrov, and Christian Andersson Naesseth. Neural flow diffusion models: Learnable forward process for improved diffusion modelling. Advances in Neural Information Processing Systems, 37: 0 73952--73985, 2024

    work page 2024

  3. [3]

    Latent dirichlet allocation

    David M Blei, Andrew Y Ng, and Michael I Jordan. Latent dirichlet allocation. Journal of Machine Learning Research, 3 0 (Jan): 0 993--1022, 2003

    work page 2003

  4. [4]

    Topic modeling using latent dirichlet allocation: A survey

    Uttam Chauhan and Apurva Shah. Topic modeling using latent dirichlet allocation: A survey. ACM Computing Surveys (CSUR), 54 0 (7): 0 1--35, 2021

    work page 2021

  5. [5]

    Density estimation using Real NVP

    Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using Real NVP . In International Conference on Learning Representations, 2017

    work page 2017

  6. [6]

    Normalizing Flow with Variational Latent Representation

    Hanze Dong, Shizhe Diao, Weizhong Zhang, and Tong Zhang. Normalizing flow with variational latent representation, 2022. URL https://arxiv.org/abs/2211.11638

    work page Pith review arXiv 2022

  7. [7]

    Neural spline flows

    Conor Durkan, Artur Bekasov, Iain Murray, and George Papamakarios. Neural spline flows. Advances in Neural Information Processing Systems, 32, 2019

    work page 2019

  8. [8]

    Mean flows for one-step generative modeling

    Zhengyang Geng, Mingyang Deng, Xingjian Bai, J Zico Kolter, and Kaiming He. Mean flows for one-step generative modeling. arXiv preprint arXiv:2505.13447, 2025

    arXiv 2025

Show all 33 references
  1. [9]

    Starflow: Scaling latent normalizing flows for high-resolution image synthesis

    Jiatao Gu, Tianrong Chen, David Berthelot, Huangjie Zheng, Yuyang Wang, Ruixiang Zhang, Laurent Dinh, Miguel Angel Bautista, Josh Susskind, and Shuangfei Zhai. Starflow: Scaling latent normalizing flows for high-resolution image synthesis. arXiv preprint arXiv:2506.06276, 2025

    2025 arXiv

  2. [10]

    beta-vae: Learning basic visual concepts with a constrained variational framework

    Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International Conference on Learning Representations, 2017

    2017

  3. [11]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33: 0 6840--6851, 2020

    2020

  4. [12]

    Semi-supervised learning with normalizing flows

    Pavel Izmailov, Polina Kirichenko, Marc Finzi, and Andrew Gordon Wilson. Semi-supervised learning with normalizing flows. In International Conference on Machine Learning, pages 4615--4630. PMLR, 2020

    2020

  5. [13]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In International Conference on Learning Representations (ICLR), 2014. URL http://arxiv.org/abs/1312.6114. Banff, Canada

    2014 arXiv

  6. [14]

    Glow: Generative flow with invertible 1x1 convolutions

    Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. Advances in Neural Information Processing Systems, 31, 2018

    2018

  7. [15]

    Improved variational inference with inverse autoregressive flow

    Durk P Kingma, Tim Salimans, Rafal Jozefowicz, Xi Chen, Ilya Sutskever, and Max Welling. Improved variational inference with inverse autoregressive flow. Advances in Neural Information Processing Systems, 29, 2016

    2016

  8. [16]

    Jet: A modern transformer-based normalizing flow

    Alexander Kolesnikov, Andr \'e Susano Pinto, and Michael Tschannen. Jet: A modern transformer-based normalizing flow. arXiv preprint arXiv:2412.15129, 2024

    2024 arXiv

  9. [17]

    On the representation of continuous functions of several variables by superpositions of continuous functions of a smaller number of variables

    Andre Nikolaevich Kolmogorov. On the representation of continuous functions of several variables by superpositions of continuous functions of a smaller number of variables. American Mathematical Society, 1961

    1961

  10. [18]

    Variational inference of disentangled latent concepts from unlabeled observations

    Abhishek Kumar, Prasanna Sattigeri, and Avinash Balakrishnan. Variational inference of disentangled latent concepts from unlabeled observations. In International Conference on Learning Representations (ICLR), 2018. URL https://arxiv.org/abs/1711.00848

    2018 arXiv

  11. [19]

    Fractalnet: Ultra-deep neural networks without residuals

    Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Fractalnet: Ultra-deep neural networks without residuals. In International Conference on Learning Representations (ICLR), 2017

    2017

  12. [20]

    Mage: Masked generative encoder to unify representation learning and image synthesis

    Tianhong Li, Huiwen Chang, Shlok Mishra, Han Zhang, Dina Katabi, and Dilip Krishnan. Mage: Masked generative encoder to unify representation learning and image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2142--2152, 2023

    2023

  13. [21]

    Fractal generative models

    Tianhong Li, Qinyi Sun, Lijie Fan, and Kaiming He. Fractal generative models. arXiv preprint arXiv:2502.17437, 2025

    2025 arXiv

  14. [22]

    Kan: Kolmogorov-arnold networks

    Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Solja c i \'c , Thomas Y Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756, 2024

    2024 arXiv

  15. [23]

    Masked autoregressive flow for density estimation

    George Papamakarios, Theo Pavlakou, and Iain Murray. Masked autoregressive flow for density estimation. Advances in Neural Information Processing Systems, 30, 2017

    2017

  16. [24]

    Normalizing flows for probabilistic modeling and inference

    George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshminarayanan. Normalizing flows for probabilistic modeling and inference. Journal of Machine Learning Research, 22 0 (57): 0 1--64, 2021

    2021

  17. [25]

    Image transformer

    Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In International Conference on Machine Learning, pages 4055--4064. PMLR, 2018

    2018

  18. [26]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn 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

  19. [27]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in Neural Information Processing Systems, 37: 0 84839--84865, 2024

    2024

  20. [28]

    Conditional image generation with pixelcnn decoders

    Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelcnn decoders. Advances in Neural Information Processing Systems, 29, 2016

    2016

  21. [29]

    Pixel recurrent neural networks

    A \"a ron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International Conference on Machine Learning, pages 1747--1756. PMLR, 2016

    2016

  22. [30]

    Generative latent flow

    Zhisheng Xiao, Qing Yan, and Yali Amit. Generative latent flow. arXiv preprint arXiv:1905.10485, 2019

    1905 arXiv

  23. [31]

    Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection

    Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. In European Conference on Computer Vision, pages 92--108. Springer, 2024

    2024

  24. [32]

    Deep learning for geophysics: Current and future trends

    Siwei Yu and Jianwei Ma. Deep learning for geophysics: Current and future trends. Reviews of Geophysics, 59: 0 e2021RG000742, 2021

    2021

  25. [33]

    Transformers without normalization

    Jiachen Zhu, Xinlei Chen, Kaiming He, Yann LeCun, and Zhuang Liu. Transformers without normalization. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 14901--14911, 2025

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.