Pith. sign in

REVIEW 5 major objections 5 minor 24 references

Improving the Generation of VAEs with High Dimensional Latent Spaces by the use of Hyperspherical Coordinates

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that rewriting a VAE's KL-type loss in hyperspherical coordinates and steering the latent angles toward zero compresses the high-dimensional latent onto a small hyperspherical island, and that this compression is what…

desk verdict Real geometric idea, but the main generation comparison samples the compressed model from a test-set-fitted vMF rather than a prior, so the headline improvement is unverified. read the letter →

arxiv 2507.15900 v1 pith:V67DGM6F submitted 2025-07-21 cs.LG

classification cs.LG
keywords variationalautoencodershypersphericalcoordinateslatentspacesparsityhigh-dimensionalstatisticsgenerativemodelingself-FIDmanifoldcompressionconcentrationofmeasure
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

High-dimensional latent spaces break generation in standard VAEs: once the latent has more than about a dozen dimensions, decoding a random latent drawn from the prior produces meaningless output even though reconstruction works. The paper locates the cause in concentration of measure, which forces standard-VAE latent vectors to lie uniformly on a hypersphere whose enormous volume makes the training data an impossibly sparse set. The proposed fix rewrites the KL-type regularizer in hyperspherical coordinates and sets the angular priors so all angles are pulled toward 0, compressing the latent manifold onto a small island of the hypersphere. With that compression, the same architecture becomes a functional generative model at 128 dimensions, and on CIFAR10 and CelebA64 the compressed VAE improves both the reconstruction MSE and the generation self-FID in the useful regime. The central claim is that latent sparsity, not the model class, is what blocks meaningful generation from high-dimensional VAEs.

What carries the argument

The central object is a KL-type loss written in hyperspherical coordinates rather than Cartesian ones. A latent vector is described by a radius $r$ and $n-1$ angles $\varphi_k$, and the loss is built from batch statistics of $\cos\varphi_k$, avoiding arccos computations, with per-coordinate priors $a_{i,j}$, $b_{i,j}$ and gains $\alpha_{i,j}$, $\beta_{i,j}$. Setting $a_{\mu,k}=1$ for every angular coordinate steers the angles away from $\pi/2$ and toward a small polar region; setting $a_{\mu,r}=\sqrt{n}$ and normalizing sampled latents to that radius keeps the samples on the sphere. Because the hyperspherical volume element carries factors $\sin^{n-2}(\varphi_1)\sin^{n-3}(\varphi_2)\cdots$, moving the angles away from the equator shrinks the latent volume dramatically in high dimensions. A vectorized Cartesian-to-hyperspherical transform and an annealing schedule for the gain $\beta$ make the reformulation practical, adding roughly 32% to training time at 200 latent dimensions.

What would settle it

Repeat the paper's comparison but draw the compressed model's generation samples from a region fixed by the training set only, or apply the same test-set-fitted directional sampling to the standard VAE; if the standard VAE's self-FID improves as much, or the compressed model's quality collapses without test-set information, the central claim fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the VAE's learned latent distribution can be reshaped by changing only the regularizer's coordinate system. In Cartesian coordinates the KL term pushes the posterior toward a high-dimensional Gaussian, whose samples concentrate on a thin equatorial band of the hypersphere of radius $\sqrt{n}$; in hyperspherical coordinates the same regularizer can instead pull the angular coordinates $\varphi_k$ toward 0 and hold the radius near $\sqrt{n}$, shrinking the occupied volume by a geometric factor. The paper reports that with this compressed loss the posterior of a 128-dimensional MNIST VAE forms a small visible island on the hypersphere and decoded random samples from the island's fitted distribution are close in quality to the training data. On CIFAR10 and CelebA64, scanning $\beta$ and latent dimension from 50 to 1000 shows that the compressed version improves on the standard VAE in both MSE and self-FID, the latter being a FID computed between random decoded samples and the reconstructed test set, over every combination of $\beta$ and latent size in the useful region, with the trade-off frontier moved inward.

Load-bearing premise

The central claim rests on accepting that new samples for the compressed model may be drawn from a distribution matched to the full test set's latent codes and that this counts as generation; if that protocol is unfair, the reported gains may only reflect re-encoding test data rather than sampling from the model's prior.

Editorial extensions

If this is right

  • At 128 latent dimensions, decoding latents drawn from the compressed MNIST model's fitted latent distribution yields digits close in quality to the training data, whereas sampling the standard VAE's prior decodes to meaningless noise.
  • On CIFAR10 and CelebA64, the compressed VAE beats the standard VAE on both reconstruction MSE and generation self-FID for every tested $\beta$ and latent size in the useful regime, so the method shifts the reconstruction-generation trade-off frontier inward.
  • Because the encoder and decoder architectures are unchanged, the comparison isolates the effect of the hyperspherical regularizer and latent-volume compression from any architectural advantage.
  • Visible class structure on the compressed 3D latent sphere suggests that the same information occupies far fewer latent configurations, which is the concrete mechanism the paper claims for reduced sparsity.
  • The only reported cost of achieving this is the coordinate transform and annealing: about 32% more training time per epoch at 200 latent dimensions.

Reading between the lines

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

  • Editorial inference: if sparsity is the mechanism, the benefit should grow with latent dimension until the transform's numerical cost dominates; a scaling study on latent sizes beyond 1000 would test this prediction directly.
  • Editorial inference: the paper's generation protocol uses a distribution fitted to the full test set's latent codes, so the practical system is a two-stage model, namely a compressed VAE plus an empirical latent distribution; applying the same protocol to the standard VAE would show how much of the gain is compression rather than the sampling protocol.
  • Editorial inference: the same coordinate-formulated regularizer could be applied to other latent-manifold models that suffer from high-dimensional sparsity, such as hierarchical or diffusion-based autoencoders, provided the pole singularities of spherical coordinates are handled explicitly.
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

5 major / 5 minor

Summary. The paper addresses the well-known failure of standard VAEs to generate meaningful data when sampling from the prior in high-dimensional latent spaces. It attributes this failure to latent sparsity caused by concentration of measure, and proposes to reparameterize the latent variables in hyperspherical coordinates in order to compress the latent manifold onto a small island of the hypersphere. The proposed training objective, given in Eqs. (5)-(7), replaces the Cartesian KL divergence with a set of squared-difference terms on the hyperspherical coordinates, introducing several hand-specified priors and gains. The main empirical claim, stated in Section IV.E, is that this 'compression VAE' improves over a standard VAE for every combination of beta and latent dimension on CIFAR10, with supporting experiments on MNIST and CelebA64. The paper also contributes a vectorized implementation of the Cartesian-to-hyperspherical conversion and visualizations of the compressed latent structure.

Significance. If the central claim were established, the volume-compression mechanism would be a simple and potentially useful addition to VAE training, and the paper's diagnosis of high-dimensional latent sparsity is a plausible and clearly presented hypothesis. The paper also has genuine strengths: it provides reproducible code for the coordinate conversion, it makes the problem of prior sampling in high-dimensional VAE latents concrete, and its qualitative latent-space visualizations (compressed island, visible clusters) are instructive. However, the quantitative evaluation as reported does not support the central claim, because the proposed model is evaluated by sampling from a test-set-fitted von Mises-Fisher distribution rather than from its prior, while the baseline is evaluated by prior sampling. This is a load-bearing fairness problem, and it is compounded by a heuristic loss derivation and a self-FID metric that compares generated samples to reconstructed test-set images. The paper therefore does not currently demonstrate that the proposed method improves generation.

major comments (5)
  1. [Section IV.D] The generation protocol for the proposed model is not a fair comparison with the baseline. The text states that 'for generating new data the latent was not randomly sampled on the whole hypersphere, but from a von Mises-Fisher distribution with the same mean and covariance as the ones empirically calculated from the latent embedding of the full test dataset.' The baseline, by contrast, is decoded from samples of N(0,I). This means the proposed model is sampled from an empirical posterior fitted to test-set latent codes, so its generated samples are near the learned data manifold by construction, regardless of whether the hyperspherical loss has any generative benefit. The reported self-FID therefore measures posterior reconstruction quality rather than prior-sampling generation quality, and the claim in Section IV.E that the compression VAE 'improves on absolute terms' is unsupported. To support the claim, the proposed model must be sampled from its actual prior, or the baseline must be evaluated with the same test-set-fitted sampling protocol, with both results reported.
  2. [Section III, Eqs. (4)-(6)] Eq. (4) is not the analytic KL divergence between the variational posterior and the prior; it is a heuristic replacement using squared differences of batch mean and batch standard deviation statistics. Eqs. (5)-(6) then introduce a large set of free parameters (alpha, beta, a, b) without any derivation from a KL divergence or any other principled variational objective. The paper acknowledges this by calling the quantity 'KLD-like', but the central methodological claim is that rewriting the latent in hyperspherical coordinates and compressing the manifold improves generation. Without a derivation of the loss, or at least a clear reframing of the method as a heuristic regularizer with justification for the parameter choices, the theoretical motivation is not established.
  3. [Section IV.E] The 'self-FID' metric compares randomly decoded samples to the reconstructed testing set, not to the original test or training images. This conflates reconstruction error with generation quality. Because the proposed model's generation protocol draws from a vMF distribution fit to test-set latent codes and decodes those draws, the decoded samples are close to test-set reconstructions by construction, making a low self-FID partly tautological. The manuscript should report standard FID against the original dataset and should compare reconstruction quality of the two models separately, so that generation quality is not conflated with reconstruction accuracy.
  4. [Conclusion, first bullet] The paper's own conclusion states that 'the improvement in generation was only evaluated for the purposes of hypothesis testing, and not as absolute performance.' This directly contradicts the abstract's claim that 'we show that this improves the generation ability of the VAE' and Section IV.E's statement that the compression VAE 'improves on absolute terms over the standard VAE.' The claims in the abstract, introduction, and Section IV.E need to be scaled back to match this stated limitation, or the experiments need to be redone with the fair protocol described in the first major comment.
  5. [Section IV.B] The loss in Eqs. (5)-(6) contains many hand-specified priors and gains, including a_mu,k = 1, a_mu,r = sqrt(n), and the 1/sqrt(k+1) scaling of alpha and beta. Section IV.B states that this scaling was chosen because it gave the best results experimentally, but no sensitivity analysis is provided. Given the number of free parameters, it is unclear whether the reported advantage is robust to reasonable changes in these hyperparameters. The paper should include ablations or report ranges of hyperparameter values over which the advantage holds.
minor comments (5)
  1. [Throughout] There are multiple typographical issues, including 'V AEs' in the abstract and elsewhere, 'sumarizes' in Section IV.E, 'fist' for 'first' in Section IV.E, and inconsistent spacing in 'V AE' and 'βV AE'.
  2. [Appendix A, Eq. (10)] In the conversion formulas, the coordinate equations list x2 twice; the second equation beginning 'x2 = r sin(phi1) sin(phi2) cos(phi3)' should almost certainly be x3. This should be corrected.
  3. [Section I.B] The statement that the latent vectors of a standard VAE are 'by construction distributed uniformly on a hypersphere' is an asymptotic statement; for finite dimension n, the angular distribution of N(0,I) is not exactly uniform. The paper should qualify this claim to avoid overstating the concentration result.
  4. [Section II] The paper says existing hyperspherical VAE methods are not suitable for comparison, but it does not report any quantitative comparison with them (e.g., S-VAE). Even if the comparison is not central, listing why these methods are different and whether any were run as baselines would make the experimental section more complete.
  5. [Appendix B] The vectorized code applies 0.001 as an additive smoothing constant in the denominator. The impact of this offset on the reported results is not discussed; a brief comment or sensitivity check would be helpful.

Circularity Check

1 steps flagged · score 7.0 of 10

Central generation claim is evaluated by sampling from a vMF fitted to test-set latent codes, not from a prior, so the comparison to a prior-sampled baseline is partly constructed.

  1. fitted input called prediction [Section IV.D, 'Improved generation when the latent manifold is compressed']
    "In this configuration, for generating new data the latent was not randomly sampled on the whole hypersphere, but from a von Mises–Fisher distribution with the same mean and covariance as the ones empirically calculated from the latent embedding of the full test dataset."

    The 'generated' samples for the compressed VAE are not drawn from the model's prior or from the learnt latent distribution over training data; they are drawn from a vMF distribution whose mean and covariance are fitted to the latent codes of the full test set. Decoding these samples therefore produces outputs that are, by construction, near the test-set latent manifold. The baseline standard VAE, in contrast, is evaluated by decoding random samples from its Gaussian prior. The paper's central empirical claim—that the compressed version 'improves on absolute terms over the standard VAE'—uses exactly this asymmetric protocol.

full rationale

The paper's main quantitative support for 'compression improves generation' is Fig. 2 and the sentence in Section IV.E stating that the compressed VAE improves over the standard VAE for every beta and latent dimension. That evidence relies on the protocol in Section IV.D, where compressed-model 'generation' is performed by sampling from a von Mises–Fisher distribution with mean and covariance empirically computed from the full test-set latent embeddings. This is a fitted posterior sampler, not a prior or model-distribution sampler, and it leaks test-set statistics into the samples. The self-FID comparison then contrasts this posterior resampling with prior sampling for the standard VAE, so the advantage is partially constructed by the evaluation design. The paper itself concedes in the conclusion that the improvement 'was only evaluated for the purposes of hypothesis testing, and not as absolute performance.' No load-bearing self-citation chain is present; the hyperspherical-coordinate rewrite and volume-compression idea have independent geometric content. The circularity is specific to the empirical generation claim, which reduces to a fit of the sampling distribution to the test data. Score 7 reflects that the central claim's experimental support is substantially undermined by this fitted-input protocol, though the underlying method is not definitionally equivalent to its inputs.

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

The method rests on a hand-built loss whose constants are either geometrically inspired (a = sqrt(n)) or tuned by experiment (alpha, beta), plus an evaluation distribution fitted to test latents. The high-dimensional concentration facts are standard, but the KLD surrogates are not derived. No new physical or mathematical entities are introduced.

free parameters (5)
  • a_mu,k (angular prior target) = 1 (all k).
    Hand-set in Section IV.D to push all angles toward 0; the reconstruction term balances it, so the effective target is data-dependent.
  • a_mu,r (radial prior target) = sqrt(n).
    Set in Eq. (9) to place latents on the sphere of radius sqrt(n); the choice is geometrical but arbitrary.
  • alpha_i,j and beta_i,j (loss gains) = Proportional to 1/sqrt(k+1), constants unspecified.
    Section IV.B states the scaling was chosen because it gave the best results experimentally; this is a free tuning choice affecting the loss.
  • vMF sampling mean and covariance for evaluation = Empirical mean and covariance of test-set latent codes.
    Section IV.D fits these to the full test set before measuring generation quality; this is parameter fitting in the evaluation itself.
  • Beta annealing schedule = sqrt(epoch) for first 100 epochs, constant after.
    The authors state the annealing was necessary to make compression work, but it is an experimental hyperparameter with no theoretical justification.
assumptions (4)
  • standard math High-dimensional Gaussian samples concentrate on the hypersphere of radius sqrt(n) and become near-uniform on it.
    Invoked in Section I.B via [6]; used to justify the claim that standard VAE latents are sparse. This is a standard concentration-of-measure result, not proved in the paper.
  • ad hoc to paper The Cartesian KL divergence can be replaced by squared differences of batch mean and batch standard deviation statistics as in Eq. (4).
    No derivation is given, and the expression is not the analytic KL between Gaussians, so the paper's loss is a surrogate.
  • ad hoc to paper A KLD-like objective in hyperspherical coordinates can be written as independent per-coordinate squared terms in cos(phi_k) and r as in Eqs. (5)-(6).
    The terms ignore coordinate dependencies and Jacobian factors; the authors explicitly avoid deriving exact formulas in Section IV.B.
  • domain assumption Angular coordinates of latent samples can be treated as uncorrelated, so setting each a_mu,k pushes samples away from the equator.
    Stated in Section III.A ('as all are uncorrelated with each other'); this assumes a factorization that the hyperspherical volume element and real encoder outputs do not guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving the Generation of VAEs with High Dimensional Latent Spaces by the use of Hyperspherical Coordinates." pith.science (2026). https://pith.science/paper/V67DGM6F

@misc{pith2026250715900,
  author       = {Pith},
  title        = {Pith review of: Improving the Generation of VAEs with High Dimensional Latent Spaces by the use of Hyperspherical Coordinates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V67DGM6F}},
  note         = {Machine review of arXiv:2507.15900}
}
read the original abstract

Variational autoencoders (VAE) encode data into lower-dimensional latent vectors before decoding those vectors back to data. Once trained, decoding a random latent vector from the prior usually does not produce meaningful data, at least when the latent space has more than a dozen dimensions. In this paper, we investigate this issue by drawing insight from high dimensional statistics: in these regimes, the latent vectors of a standard VAE are by construction distributed uniformly on a hypersphere. We propose to formulate the latent variables of a VAE using hyperspherical coordinates, which allows compressing the latent vectors towards an island on the hypersphere, thereby reducing the latent sparsity and we show that this improves the generation ability of the VAE. We propose a new parameterization of the latent space with limited computational overhead.

Figures

Figures reproduced from arXiv: 2507.15900 by the authors.

Figure 1
Figure 1. Comparison using MNIST between the standard [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Effect of latent dimension and β on the trade off between reconstruction and generation on CIFAR10. Each curve represents a VAE for a given β while spanning the number of latent dimensions from low (50, bottom right endpoints) to high (1000 top left corner endpoints). The standard VAEs are shown using dashed lines, whereas the compressed versions are shown using solid lines. We excluded from our discussion the regim… view at source ↗
Figure 3
Figure 3. Effect of latent dimension and β on the trade-off between reconstruction and generation on CelebA64 (as in CIFAR10, solid lines closer to the bottom left corner than the dashed lines). [8] L. P. Cinelli, M. A. Marins, E. A. B. da Silva, and S. L. Netto, Variational methods for machine learning with applications to deep networks. Springer International Publishing, 5 2021. [9] P. A. Jonathan Ho Ajay Jain, “Denoising D… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 19 canonical work pages

  1. [1]

    Auto-Encoding Variational Bayes,

    M. W. Diederik P Kingma, “Auto-Encoding Variational Bayes,” International Conference on Learning Representations , vol. 2, 2014. [Online]. Available: https://arxiv.org/abs/1312.6114

  2. [2]

    An Introduction to Variational Autoencoders,

    D. P. Kingma and M. Welling, “An Introduction to Variational Autoencoders,” Foundations and Trends® in Machine Learning , vol. 12, no. 4, pp. 307–392, 2019. [Online]. Available: https: //www.nowpublishers.com/article/Details/MAL-056

  3. [3]

    Variational Autoencoders Pursue PCA Directions (by Accident),

    M. Rol ´ınek, D. Zietlow, and G. Martius, “Variational Autoencoders Pursue PCA Directions (by Accident),” Tech. Rep., 2019

  4. [4]

    Why do Variational Autoencoders Really Promote Disentanglement?

    P. Bhowal, A. Soni, and S. Rambhatla, “Why do Variational Autoencoders Really Promote Disentanglement?” Tech. Rep., 2024. [Online]. Available: https://github.com/criticalml-uw/

  5. [5]

    β-V AE: Learning Basic Visual Concepts with a Constrained Variational Framework,

    I. Higgins, L. Matthey, A. Pal, C. Burgess, X. Glorot, M. Botvinick, S. Mohamed, A. Lerchner, and G. Deepmind, “ β-V AE: Learning Basic Visual Concepts with a Constrained Variational Framework,” Tech. Rep., 2017

  6. [6]

    Vershynin, High-Dimensional Probability

    R. Vershynin, High-Dimensional Probability . Cambridge University Press, 4 2018. [Online]. Available: https://doi.org/10.1017/ 9781108231596

  7. [7]

    Concentration of measure,

    M. J. Wainwright, “Concentration of measure,” in High-Dimensional Statistics . Cambridge University Press, 2 2019, ch. 3, pp. 58–97. [Online]. Avail- able: https://www.cambridge.org/core/books/highdimensional-statistics/ concentration-of-measure/A649A3B05DC79C2B10BF1C80CC6F5F10 Fig. 3. Effect of latent dimension and β on the trade-off between reconstructi...

  8. [8]

    L. P. Cinelli, M. A. Marins, E. A. B. da Silva, and S. L. Netto, Variational methods for machine learning with applications to deep networks. Springer International Publishing, 5 2021

Show all 24 references
  1. [9]

    Denoising Diffusion Probabilistic Models,

    P. A. Jonathan Ho Ajay Jain, “Denoising Diffusion Probabilistic Models,” arXiv preprint , 2020. [Online]. Available: https://arxiv.org/ abs/2006.11239

  2. [10]

    Interpreting the Curse of Dimensionality from Distance Concentration and Manifold Effect,

    D. Peng, Z. Gui, and H. Wu, “Interpreting the Curse of Dimensionality from Distance Concentration and Manifold Effect,” Tech. Rep., 2023

  3. [11]

    Hyperspherical Variational Auto-Encoders,

    T. R. Davidson, L. Falorsi, N. De Cao, T. Kipf, and J. M. Tomczak, “Hyperspherical Variational Auto-Encoders,” Tech. Rep.,

  4. [12]

    Spherical Sliced-Wasserstein,

    C. Bonet, P. Berg, N. Courty, F. Septier, L. Drumetz, and M.-T. Pham, “Spherical Sliced-Wasserstein,” 6 2022. [Online]. Available: http://arxiv.org/abs/2206.08780

  5. [13]

    Sliced-Wasserstein Autoencoder: An Embarrassingly Simple Generative Model,

    C. E. M. G. K. R. Soheil Kolouri Phillip E. Pope, “Sliced-Wasserstein Autoencoder: An Embarrassingly Simple Generative Model,” 2020. [Online]. Available: https://arxiv.org/abs/1804.01947

  6. [14]

    Rotating Features for Object Discovery,

    S. L ¨owe, P. Lippe, F. Locatello, and M. Welling, “Rotating Features for Object Discovery,” 6 2023. [Online]. Available: http: //arxiv.org/abs/2306.00600

  7. [15]

    VICReg: Variance-Invariance- Covariance Regularization for Self-Supervised Learning,

    A. Bardes, J. Ponce, and Y . LeCun, “VICReg: Variance-Invariance- Covariance Regularization for Self-Supervised Learning,” 5 2021. [Online]. Available: http://arxiv.org/abs/2105.04906

  8. [16]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” Tech. Rep., 2015. [Online]. Available: http://image-net.org/challenges/LSVRC/2015/

  9. [17]

    Cyclical Annealing Schedule: A Simple Approach to Mitigating KL Vanishing,

    H. Fu, C. Li, X. Liu, J. Gao, A. Celikyilmaz, and L. Carin, “Cyclical Annealing Schedule: A Simple Approach to Mitigating KL Vanishing,” 3 2019. [Online]. Available: http://arxiv.org/abs/1903.10145

  10. [18]

    Gradient-based learning applied to document recogni- tion

    Y .LeCun et al, “Gradient-based learning applied to document recogni- tion.” Proceedings of the IEEE , vol. 86, no. 11, pp. 2278–2324, 11 1998

  11. [19]

    GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium,” 6 2017. [Online]. Available: http: //arxiv.org/abs/1706.08500

  12. [20]

    Learning Multiple Layers of Features from Tiny Im- ages,

    A. Krizhevsky, “Learning Multiple Layers of Features from Tiny Im- ages,” Tech. Rep., 2009

  13. [21]

    Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models,

    G. Stein, J. C. Cresswell, R. Hosseinzadeh, Y . Sui, B. Leigh Ross, V . Villecroze, Z. Liu, A. L. Caterini, J. T. Eric Taylor, and G. Loaiza-Ganem, “Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models,” in 37th Conference on Neu...

  14. [22]

    Deep Learning Face Attributes in the Wild,

    Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep Learning Face Attributes in the Wild,” Proceedings of International Conference on Computer Vision (ICCV), 12 2015

  15. [2018]

    Available: https://github.com/nicola-decao/s-vae

    [Online]. Available: https://github.com/nicola-decao/s-vae

  16. [2023]

    Available: https://github.com/layer6ai-labs/dgm-eval

    [Online]. Available: https://github.com/layer6ai-labs/dgm-eval

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.