Pith. sign in

REVIEW 5 major objections 5 minor 28 references

Wavelet-based Variational Autoencoders for High-Resolution Image Generation

T0 review · 5 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read Using Haar wavelet coefficients as the latent representation of a VAE, with L1 sparsity regularization, reduces blur in reconstructions compared with a Gaussian latent VAE in reported single-run experiments.

desk verdict Interesting wavelet-VAE combination, but Eq. (22) is not a valid ELBO and no generative sampling path is defined, so the central claim of high-resolution image generation is unsupported; the experiments only show reconstruction. read the letter →

arxiv 2504.13214 v1 pith:24NP4JIY submitted 2025-04-16 cs.CV eess.IV

classification cs.CVeess.IV
keywords latentvaeswavelet-basedautoencoderscoefficientsconventionaldetailsgenerative
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

Variational autoencoders (VAEs) are models that compress an image into a small set of hidden numbers, then try to rebuild the image from those numbers. They usually make blurry pictures because the hidden numbers are assumed to follow a simple bell-shaped curve and because pixel-by-pixel losses average out fine details. This paper proposes storing the hidden information as wavelet coefficients, which are numbers that describe both where features are and how sharp they are, like a built-in multi-scale edge detector. The encoder outputs these coefficients, and the decoder is a fixed inverse wavelet transform. To keep the model somewhat random, the authors add Gaussian noise to the coefficients, and that noise level is learned during training. They replace the usual KL-divergence penalty with a simple L1 penalty that pushes most detail coefficients toward zero, relying on the natural sparsity of wavelet coefficients. The paper reports that on CIFAR-10 at 128x128 resolution, the wavelet version has a lower reconstruction error (0.038 versus 0.045), a higher SSIM (0.79 versus 0.70), and a lower FID (28.1 versus 32.5) than a conventional VAE. The authors claim sharper reconstructions and better high-frequency detail. However, the paper does not describe the architecture, training details, baseline VAE, or any generated samples from the prior, and it drops the KL term without showing that the objective remains a valid variational bound. The CelebA-HQ results promised in the abstract are not shown in the tables.
Extended reading notes

Core claim

The paper's central assertion is that replacing the Gaussian latent space with multi-scale Haar wavelet coefficients, a learnable noise scale, and L1 sparsity regularization substantially reduces blur and recovers higher-resolution details compared to conventional VAEs. The abstract states: 'Our experimental evaluation on CIFAR-10 and other high-resolution datasets demonstrates that the Wavelet-VAE improves visual fidelity and recovers higher-resolution details compared to conventional VAEs.' If true, the model at least offers a sharper reconstruction alternative, though the evidence is limited to CIFAR-10 reconstruction metrics.

Load-bearing premise

The weakest load-bearing premise is that the objective in Eq. (22), which discards the KL divergence and replaces it with an L1 penalty on deterministic encoder outputs, is a valid variational objective, and that the learnable noise scale s will not collapse to zero during optimization. This premise enters at Eq. (22) and is tested only indirectly in Section 4.5. If s collapses to zero, the model becomes a deterministic sparse autoencoder with a fixed linear decoder, not a stochastic generator. The paper provides no theoretical or empirical evidence that s remains positive, and with a fixed reconstruction likelihood the gradient of the loss with respect to s points toward zero. Thus the 'variational' and 'stochastic generation' aspects of the central claim rest on an unvalidated assumption.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes Wavelet-VAE, a variational autoencoder variant whose latent space is composed of multi-scale Haar wavelet coefficients. The encoder outputs deterministic wavelet coefficients, to which a learnable noise scale s is added; the decoder is a fixed inverse discrete wavelet transform. The training objective in Eq. (22) drops the standard KL term and instead applies an L1 sparsity penalty to the encoder's detail coefficients. The authors claim that this approach reduces blur and recovers higher-resolution details, and they report reconstruction loss, SSIM, and FID for a conventional VAE and the Wavelet-VAE on CIFAR-10 upscaled to 128x128, while also stating that CelebA-HQ at 128x128 and 256x256 would be used.

Significance. If the model were a valid generative model with the reported gains, the wavelet-based structured latent representation would be a potentially useful contribution to VAE research, and the interpretability discussion in Section 5 is a genuine strength. The learnable-noise idea is also interesting, although its current formulation is problematic. However, the paper provides no code, no machine-checked proofs, and the central claims rest on an objective that is not a valid ELBO and on an experimental section that does not match the abstract's dataset claims. As presented, the method is a regularized sparse autoencoder with a fixed linear decoder, not a generative model; the reported FID and SSIM improvements are therefore not evidence for improved generation.

major comments (5)
  1. [3.3.2, Eq. (22)] Equation (22) is not a valid ELBO. The standard objective in Eq. (14) contains the KL divergence D_KL(q_phi(c~|x) || p(c)), which is necessary for the objective to lower-bound the log-likelihood and to tie the posterior to a sampling prior. In Section 3.3.2, this term is replaced by an L1 penalty on the deterministic encoder outputs c_NN, with no derivation that this regularizer is a bound on the KL divergence or a surrogate that preserves the variational interpretation. A model trained with Eq. (22) is therefore a regularized autoencoder, not a variational autoencoder, and the paper's central claim of a 'wavelet-based variational autoencoder' is unsupported.
  2. [3.2.2, 4] No generative sampling path is defined. The decoder is the fixed inverse discrete wavelet transform (IDWT) applied to noisy coefficients c~ = c_NN + s*epsilon, and training minimizes a reconstruction loss on input images. At generation time, one would need to sample coefficients from a prior, but the paper never specifies how to sample from a distribution that matches the trained coefficient statistics; simply invoking the Laplacian prior in Eq. (20) is insufficient because the encoder was trained to produce DWT-like coefficients of real images and the fixed decoder was never trained on prior samples. The FID values in Table 1 cannot be interpreted as generative quality if they were computed on reconstructions, and the paper does not describe a protocol for measuring generation quality.
  3. [3.2.2, 4.5] The learnable noise scale s is at risk of collapsing to zero. For a fixed encoder output c_NN and a reconstruction loss such as MSE, the expected loss over epsilon is minimized at s=0 because the additive noise is zero-mean and only inflates the reconstruction error; the gradient expression in Section 3.2.2 does not contain any mechanism (e.g., a KL term or variance constraint) that would counteract this collapse. Section 4.5 reports only that a learnable s outperforms a fixed s=0.01, but it does not report the learned value of s at convergence or demonstrate that the model remains stochastic. If s collapses, the model is a deterministic sparse autoencoder, contradicting the claim of stochastic generation.
  4. [4.1, 4.3] The experimental section does not support the datasets claimed in the abstract and introduction. Section 4.1 states that CelebA-HQ at 128x128 and 256x256 will be evaluated, but Table 1 and the quantitative discussion in Section 4.3 report results only for CIFAR-10 upscaled to 128x128. No CelebA-HQ quantitative or qualitative results appear anywhere in the manuscript. The abstract's claim of demonstrating improvements on 'CIFAR-10 and other high-resolution datasets' is therefore unsubstantiated.
  5. [4.2, Table 1] The comparison is not reproducible because the baseline VAE architecture, latent dimension, wavelet decomposition level L, and training hyperparameters are not specified, and no error bars or multiple-seed statistics are reported. With only a single run per model, the reported differences in reconstruction loss (0.045 vs 0.038) and SSIM (0.70 vs 0.79) may reflect run-to-run variability rather than a real improvement. A clearly specified baseline and confidence intervals (or a much larger experimental study) are needed to support the claimed gains.
minor comments (5)
  1. [1.3] The cross-reference 'Section??' in the contributions overview is unresolved; it should be replaced with the actual section number introducing the architecture.
  2. [References] Entries [15] and [16] both cite arXiv:2504.03821, which appears to be an erroneous duplicate identifier; the two cited works have distinct titles and should have distinct identifiers.
  3. [3.2.2] Notation for the encoder output is inconsistent: 'c_{i,NN}', 'cNN', and 'c_i,NN(x;phi)' are used interchangeably; please unify the notation throughout Section 3.
  4. [Table 1, Sec. 4.2] The 'Recon Loss' column in Table 1 is not defined; Section 4.2 mentions both MSE and BCE without stating which one is reported in the table.
  5. [4.5] The ablation states that a fixed noise scale gives 'approximately 8% higher MSE' but does not report the actual MSE values, the exact fixed scale used beyond an example of 0.01, or the converged value of the learnable s.
Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No invented entities are introduced; the model relies on standard wavelet transforms and network components. The main burden is the unvalidated variational objective, which assumes that dropping the KL term and using an L1 penalty still yields a useful generative model. The free parameters (lambda, s, likelihood variance, architectural choices) are fitted or chosen without sufficient reporting, limiting the evidential weight of the empirical comparison.

free parameters (4)
  • Sparsity weight lambda = not reported
    Hyperparameter balancing reconstruction loss and L1 penalty in Eqs. (21)-(22); no value or tuning procedure is given in the paper.
  • Noise scale s = not reported
    Learnable parameter in Eq. (12) intended to preserve stochasticity; no final learned value is reported, and the objective does not prevent it from collapsing to zero.
  • Reconstruction likelihood variance = not specified
    The reconstruction log-likelihood term in Eq. (22) is not fully specified; if treated as Gaussian with fixed variance, that variance is an unstated free parameter affecting the loss scale.
  • Wavelet decomposition level L and latent dimensions = not reported
    Depth of Haar decomposition and latent dimensionality are not specified; these architectural choices affect the comparison and are not justified.
assumptions (4)
  • standard math DWT/IDWT with Haar filters achieves perfect reconstruction: IDWT(DWT(x)) = x.
    Used in Section 3.1.3 to justify the fixed decoder design; this is a standard property of orthogonal wavelets.
  • domain assumption Natural image wavelet coefficients are sparse, so L1 regularization on detail coefficients is an appropriate substitute for the KL term.
    Invoked in Section 3.3.2 to replace D_KL with an L1 penalty; empirical sparsity of wavelet coefficients in natural images is assumed, not verified in the model's actual learned latent space.
  • ad hoc to paper Replacing the KL divergence with an L1 penalty on encoder outputs yields a valid variational objective that supports sampling from a prior at generation time.
    Eq. (22) drops the KL term entirely; the paper does not derive a lower bound or specify the prior used for generation, so this assumption is asserted rather than proved.
  • domain assumption A fixed linear IDWT decoder has enough capacity to represent the image distribution given noisy wavelet coefficients.
    The decoder is a fixed inverse transform (Section 3.2.2); the paper does not analyze whether this restricts the model's generative ability or capacity to model complex image distributions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Wavelet-based Variational Autoencoders for High-Resolution Image Generation." pith.science (2026). https://pith.science/paper/24NP4JIY

@misc{pith2026250413214,
  author       = {Pith},
  title        = {Pith review of: Wavelet-based Variational Autoencoders for High-Resolution Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/24NP4JIY}},
  note         = {Machine review of arXiv:2504.13214}
}
read the original abstract

Variational Autoencoders (VAEs) are powerful generative models capable of learning compact latent representations. However, conventional VAEs often generate relatively blurry images due to their assumption of an isotropic Gaussian latent space and constraints in capturing high-frequency details. In this paper, we explore a novel wavelet-based approach (Wavelet-VAE) in which the latent space is constructed using multi-scale Haar wavelet coefficients. We propose a comprehensive method to encode the image features into multi-scale detail and approximation coefficients and introduce a learnable noise parameter to maintain stochasticity. We thoroughly discuss how to reformulate the reparameterization trick, address the KL divergence term, and integrate wavelet sparsity principles into the training objective. Our experimental evaluation on CIFAR-10 and other high-resolution datasets demonstrates that the Wavelet-VAE improves visual fidelity and recovers higher-resolution details compared to conventional VAEs. We conclude with a discussion of advantages, potential limitations, and future research directions for wavelet-based generative modeling.

Figures

Figures reproduced from arXiv: 2504.13214 by the authors.

Figure 1
Figure 1. Qualitative comparisons between baseline VAE (top row) and Wavelet-VAE (bottom row). [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Heatmap of Haar wavelet coefficients extracted from the latent space of a Wavelet-VAE trained [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 26 canonical work pages

  1. [16]

    A hybrid wavelet-fourier method for next-generation con- ditional diffusion models

    Andrew Kiruluta and Andreas Lemos. A hybrid wavelet-fourier method for next-generation con- ditional diffusion models. arXiv:2504.03821, 2025

  2. [1]

    Bowman, Luke Vilnis, Oriol Vinyals, Andrew M

    Samuel R. Bowman, Luke Vilnis, Oriol Vinyals, Andrew M. Dai, Rafal Jozefowicz, and Samy Bengio. Generating sentences from a continuous space. Conference on Computational Natural Language Learning (CoNLL), 2016

  3. [2]

    Importance-weighted autoencoders

    Yuri Burda, Roger Grosse, and Ruslan Salakhutdinov. Importance-weighted autoencoders. Inter- national Conference on Learning Representations (ICLR), 2016

  4. [3]

    R. Chen, X. Zhang, and J. Li. Wavelet transform network for image denoising. IEEE Transactions on Image Processing, vol. 30, pp. 2340–2352, 2021

  5. [4]

    Very deep vaes generalize autoregressive models and can outperform them on im- ages

    Rewon Child. Very deep vaes generalize autoregressive models and can outperform them on im- ages. International Conference on Learning Representations (ICLR), 2021

  6. [5]

    Denton, Soumith Chintala, Arthur Szlam, and Rob Fergus

    Emily L. Denton, Soumith Chintala, Arthur Szlam, and Rob Fergus. Deep generative image mod- els using a laplacian pyramid of adversarial networks. Neural Information Processing Systems (NeurIPS), 2015

  7. [6]

    Multi-Anchor Active Domain Adaptation for Semantic Segmentation

    Z. Fu and R. K. Ward. Wavelet-based generative adversarial networks. arXiv preprint arXiv:2108.08012, 2021

  8. [7]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Neural Information Processing Systems (NeurIPS), 2014. 13

Show all 28 references
  1. [8]

    Draw: A recurrent neural network for image generation

    Karol Gregor, Ivo Danihelka, Alex Graves, Danilo Rezende, and Daan Wierstra. Draw: A recurrent neural network for image generation. International Conference on Machine Learning (ICML), 2015

  2. [9]

    Pixelvae: A latent variable model for natural images

    Ishaan Gulrajani, Kundan Kumar Agrawal, François Duval, and Aaron Courville. Pixelvae: A latent variable model for natural images. International Conference on Learning Representations (ICLR), 2017

  3. [10]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Neural Infor- mation Processing Systems (NeurIPS), 2017

  4. [11]

    Progressive growing of gans for improved quality, stability, and variation

    Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. International Conference on Learning Representations (ICLR), 2018

  5. [12]

    An introduction to variational methods for graphical models

    Diederik Kingma and Max Welling. An introduction to variational methods for graphical models. Foundations and Trends in Machine Learning, vol. 12, no. 3, 2019

  6. [13]

    Kingma, Tim Salimans, and Max Welling

    Diederik P. Kingma, Tim Salimans, and Max Welling. Improved variational inference with inverse autoregressive flow. Neural Information Processing Systems (NeurIPS), 2016

  7. [14]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. International Conference on Learning Representations (ICLR), 2014

  8. [17]

    Alex Krizhevsky and Geoffrey E. Hinton. Learning multiple layers of features from tiny images. Technical Report, University of Toronto, 2009

  9. [18]

    Kulkarni, William F

    Tejas D. Kulkarni, William F. Whitney, Pushmeet Kohli, and Joshua B. Tenenbaum. Deep convo- lutional inverse graphics network. Neural Information Processing Systems (NeurIPS), 2015

  10. [19]

    C. Liu, F. Zhang, and Y . Wang. Wavelet-based convolutional neural networks for texture classifi- cation. Neural Computing and Applications, vol. 32, no. 9, 2020

  11. [20]

    A wavelet tour of signal processing

    Stephane Mallat. A wavelet tour of signal processing. Academic Press, 1999

  12. [21]

    Patil and S

    V . Patil and S. Patil. Use of wavelet transform in convolution neural network. IEEE International Conference on Electronics, Computing and Communication Technologies, 2020

  13. [22]

    Generating diverse high-fidelity images with vq-vae-2

    Ali Razavi, Aäron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Neural Information Processing Systems (NeurIPS), 2019

  14. [23]

    Variational inference with normalizing flows

    Danilo Jimenez Rezende and Shakir Mohamed. Variational inference with normalizing flows. International Conference on Machine Learning (ICML), 2015

  15. [24]

    Stochastic backpropagation and approximate inference in deep generative models

    Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. International Conference on Machine Learning (ICML), 2014

  16. [25]

    Ladder varia- tional autoencoders

    Casper Kaae Sønderby, Tapani Raiko, Lars Maaløe, Claus Svarer, and Ole Winther. Ladder varia- tional autoencoders. Neural Information Processing Systems (NeurIPS), 2016. 14

  17. [26]

    Tomczak and Max Welling

    Jakub M. Tomczak and Max Welling. Vae with a vampprior. Artificial Intelligence and Statistics (AISTATS), 2018

  18. [27]

    Nvae: A deep hierarchical variational autoencoder

    Arash Vahdat and Jan Kautz. Nvae: A deep hierarchical variational autoencoder. Neural Informa- tion Processing Systems (NeurIPS), 2020

  19. [28]

    Bovik, Hamid R

    Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004

  20. [29]

    Attribute2image: Conditional image generation from visual attributes

    Xinchen Yan, Jimei Yang, Kihyuk Sohn, Hong Lee, and Dahua Lin. Attribute2image: Conditional image generation from visual attributes. European Conference on Computer Vision (ECCV), 2016. 15

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.