Pith. sign in

REVIEW 4 major objections 5 minor 22 references

Decafs: Disentangled Conditional adversarial Flows

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A pretrained flow can be made conditionally controllable and factor-disentangled by adding a non-invertible Lie-group generator aligned to its latent space.

desk verdict Worth a look for the idea of aligning a non-invertible Lie-group generator to a pretrained flow, but the DCI/FVM scores are circular and the generator's map into the flow latent is never specified. read the letter →

arxiv 2607.18755 v1 pith:J2B75WB2 submitted 2026-07-21 cs.LG

Anirudh jain , Sakshi Varshney , Samuel Kaski , Vikas Garg This is my paper
classification cs.LG
keywords disentanglementnormalizingflowsconditionalgenerationLiegroupscontrolledmoleculargraphadversariallatentalignment
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

Decaf is attempting to establish a decoupling result: to condition a normalizing flow on labels or molecular properties, you do not need to retrain the flow or expand its latent dimension. Instead, you train a small, non-invertible generator that produces a latent code with a special structure—a Lie-group disentangled part, where each condition maps to a scalar coordinate along a learned matrix basis, plus a coupled part for residual correlations—and you align that code to the already-trained flow's latent space with a Wasserstein adversarial loss and a predictor loss. The paper argues that this alignment makes the conditional distributions match, so sampling from the generator and inverting the flow yields controlled samples. The payoff, if true, is interpretable and controllable generation at roughly 10% parameter cost, with strong reported results on images and molecules.

What carries the argument

The central object is the Lie-group conditional generator. It splits the latent code into a disentangled block, zd = exp(Σ ti Ai), where the {Ai} are learned matrices forming a Lie algebra basis and ti = mi(ci) are condition-dependent scalar coordinates, and a coupled block zc = g(c, ε) that captures residual correlations. The exponential map is what carries the argument: it turns scalar coordinate changes into group actions on the latent manifold, so varying ti moves the generated sample along a single generative factor. A combination block h(zc, zd) produces the final ẑ, and a Wasserstein adversarial loss together with a predictor loss aligns ẑ to the flow's latent space; the paper's key s

What would settle it

Run the trained Decaf on dSprites, fix the noise and all but one Lie coordinate ti, and decode through the inverse flow; then run the same linear classifier used by the FactorVAE metric on the decoded images. If changing ti moves more than one generative factor, or if the class posterior of the flow latent differs measurably from the generator's for the same condition, the claimed perfect disentanglement and the conditional-matching step are not reproduced.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a flow's latent space can be disentangled post hoc: given a pretrained invertible flow, Decaf trains an irreversible generator G(c, ε) whose output ẑ is aligned to the flow latent z. The generator's disentangled branch computes zd = exp(Σ ti Ai), where each ti = mi(ci) is produced by a network reading one condition label and {Ai} is a learned set of Lie algebra basis matrices; the coupled branch zc = g(c, ε) handles entangled structure, and a combination block h merges them. Alignment is enforced by a Wasserstein critic that treats flow latents as real and generator outputs as fake, together with a predictor that forces both to give the same cl

Load-bearing premise

The claim rests on the unstated correspondence between the matrix-valued Lie-group exponential and the flow's Euclidean multi-scale latent space: if the disentangled output cannot be placed in a bijective, dimension-matched mapping to the space where the flow operates, then the coordinates ti are not well-defined controls and the alignment losses are just fitting a black-box combination.

Editorial extensions

If this is right

  • Any pretrained flow can be upgraded to conditional generation by training the small add-on generator and predictor, with no retraining of the flow and no expansion of its latent dimension.
  • Users get factor-level control: on dSprites the paper shows one Lie coordinate changes shape, another size, another rotation, so a single generative factor can be moved while others stay fixed.
  • The same alignment recipe is claimed to transfer to diffusion models and flow-matching models, because it only needs access to the latent or noise space, not the flow's invertibility.
  • In molecular generation, a single forward pass through the generator can jointly target multiple properties (QED, logP, SA score), unlike prior methods that maximize one property at a time.
  • Perfect classifier-based disentanglement scores (FVM, DCI) become attainable on simple factorized datasets when the generator's Lie basis is one-to-one with the generative factors.

Reading between the lines

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

  • The paper never specifies how the matrix-valued exponential zd is reshaped or dimension-matched into the flow's multi-scale latent-space vector; if the combination block h is doing the alignment, the disentangled coordinates ti may not be the actual controls, and the 10% overhead claim understates the effective capacity used.
  • If the alignment is genuinely distribution-preserving, it suggests a general recipe for interpretability: disentangling a generative model's latent space by training a structured external generator, rather than regularizing the model itself—this could be tested by applying the same add-on to a diffusion model.
  • The perfect FVM/DCI scores are on dSprites, where factors are independent by construction; on real-world molecular properties the coupled block is essential, so the honest target is partial disentanglement, and the reported distributions show the disentangled/coupled trade-off.
  • A direct test of the mechanism: use the pretrained predictor to classify latent traversals from a single ti sweep; if the classifier detects changes in non-target factors, the factor isolation is not actually achieved.
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 / 5 minor

Summary. The paper proposes Decaf, a framework for conditional generation with normalizing flows. A pretrained unconditional flow is left unchanged, and a non-invertible generator G(c, ε) produces latent codes that are aligned to the flow latent space via a Wasserstein GAN and a predictor loss. The generator splits its latent code into a coupled part zc and a disentangled part zd = exp(Σ_i t_i A_i), where each t_i = m_i(c_i) is produced by a network from a single condition coordinate and the A_i are learned Lie-basis matrices. The paper reports lower FID than StyleGAN on MNIST, perfect DCI/FVM scores on dSprites, and conditional molecular generation on QM9, ZINC, and MOSES with property-conditioned distribution shifts.

Significance. If the mechanism were fully specified and validated, the core idea would be useful: aligning a lightweight conditional generator with a pretrained unconditional flow could enable controllable generation without modifying the flow architecture, and the reported ~10% parameter overhead is appealing. The Lie-group construction is a plausible inductive bias for factor-level control. The molecular experiments address an application domain of real interest. However, the manuscript as written does not establish the central claim of disentangled flow-latent generation: the key construction is underspecified, the mathematical derivation in Sec. 3.1 is garbled, and the reported perfect disentanglement scores appear to certify the supervised input decomposition rather than the flow latent space.

major comments (4)
  1. [§3.1, Eq. (4)] The change-of-variables identity is written incorrectly. The flow objective should be max_θ E_{x∼pdata}[log pθ(x)] = E_{x∼pdata}[log pθ(z=fθ(x)) + log|det ∂fθ/∂x|], with z deterministically equal to fθ(x). The displayed expression instead has E_{z∼pθ} inside an objective over pθ(x), which is not a valid objective and does not define the flow's log-likelihood. The same issue appears in the L_flow equation in §4.1, where the expectation is over z∼p(z) rather than over data x. Because this identity is the mathematical foundation for the alignment argument, it needs to be corrected and written consistently.
  2. [§4.1, Eq. (1)–(2); Tables 5–7] The central construction is underspecified. zd = exp(Σ_i t_i A_i) is a matrix exponential with A_i ∈ R^{|A|×|A|}, so zd is a |A|×|A| matrix, while the flow latent z is a Euclidean vector or multi-scale tensor (e.g., 4×7×7 after the splits in Table 3). The paper never defines the dimensionality of z, the projection from the matrix-valued zd to the flow latent space, or the combination block h(zc, zd). Without a dimension-matched, bijective (or at least well-defined) correspondence between the Lie-group output and the actual flow latent space, the disentangled coordinates t_i have no well-defined effect on generated samples and the alignment objective is vacuous. This must be specified and justified.
  3. [§5.1, Table 1b] The reported perfect DCI=1.00 and FVM=1.00 on dSprites are likely circular. The disentangled block defines t_i = m_i(c_i), where c_i is the one-hot encoding of factor i alone, so the coordinates t_i are disentangled by construction. The manuscript does not state which representation is fed to the DCI/FVM evaluators. If the metrics are computed on t_i or zd, they merely certify the supervised input decomposition and provide no evidence about the flow latent space actually used by F^{-1}. The correct protocol is to evaluate disentanglement on z = fθ(x) or on ẑ = G(c, ε), and to report the standard factor-prediction metrics on those codes.
  4. [§5.2, Table 8] The claim that non-target properties show 'minimal variations' is not supported by the reported numbers. For ZINC, unconditional logP is 2.45±1.44 but QED=0.5 gives logP=3.29±1.07; unconditional SA Score is 3.06±0.84 but QED=0.5 gives SA=3.60±1.13; similar shifts appear in other conditions. If these shifts are intended to reflect the coupled block, the paper needs quantitative disentanglement/coupling metrics for molecular properties rather than visual inspection of histograms. As written, the evidence for disentangled molecular control is weak.
minor comments (5)
  1. [§4.1, Algorithm 1] Algorithm 1 is self-contradictory: line 1 says 'Pre-trained Flow F', but lines 8 and 24 update the flow parameters θ via L_flow. Please clarify whether the flow is frozen or fine-tuned, and if frozen, what role L_flow plays.
  2. [Eq. (3), Algorithm 1] The WGAN objective in Eq. (3) includes a gradient-penalty term λ_GP, but Algorithm 1 line 19 only lists −D_φ(z)+D_φ(ẑ) and the optimizer update for φ does not mention the gradient penalty. Please state where λ_GP is applied.
  3. [§4.1] The text refers to 'Algorithm ??' and 'Algorithm ??' without resolving the cross-references. Also Section 3.2 is a one-sentence placeholder.
  4. [Global] Typos and notation inconsistencies: 'floxible' for 'flexible'; 'irreversible generator' presumably means 'non-invertible generator'; L_Comm = (1/T)Σ(A_iA_j−A_jA_i) in Algorithm 1 lacks a norm or absolute value and is not an obvious penalty; L_Hess = (1/T)Σ(A_iA_j) is not described in the text.
  5. [Appendix 10] Figure 10 has the placeholder caption 'Caption'; Appendix 10.1 explains DCI/FVM/SAP/MIG but does not state the representation used for Decaf in Table 1b — this is essential and should be moved to the main text.

Circularity Check

1 steps flagged · score 6.0 of 10

Perfect DCI/FVM scores merely restate the supervised construction t_i=m_i(c_i); the main disentanglement evidence is circular, though molecular benchmarks are external.

  1. self definitional [Sec. 4.1 (Lie generator definition) + Sec. 5.1 / Table 1b]
    "zd = exp( kX i=1 tiAi) where each ti = mi(ci) is obtained via an independent neural network mi that takes the corresponding condition label as input. ... Decaf achieves a perfect score for classifier-based metrics FVM and DCI as the Lie-group generator matches each generative factor to one of the Lie group basis."

    FVM and DCI are classifier-based disentanglement metrics: they train a linear classifier / random forests to predict the ground-truth generative factors from the latent representation. In Decaf, the only explicitly disentangled part of the representation, zd, is built from those same ground-truth factors by feeding one-hot c_i into per-factor networks t_i=m_i(c_i). Hence the factor labels are inputs to the scored representation, not properties discovered in the flow latent space. A near-perfect DCI/FVM is forced by the supervised input decomposition; the paper's own explanation credits the score to this matching. The paper never reports scoring the actual flow latent z=f_theta(x) or ablating the injected t_i channels, so the perfect scores do not validate disentanglement of the aligned flo

full rationale

The most load-bearing empirical claim, perfect DCI and FVM on dSprites, is circular in the sense defined: the representation being certified is constructed from the very one-hot factor labels that the metrics try to recover. DCI and FVM use classifiers/random forests to predict generative factors from latent codes; Decaf's disentangled block defines t_i=m_i(c_i) with a separate network per factor and forms zd=exp(Σ t_i A_i), so the labels are injected into the representation before the metric is computed. The paper even states the perfect score is because the generator matches each factor to a Lie basis, confirming the metric certifies the construction rather than a learned disentanglement of the flow latent. The molecular conditional-generation results are not similarly circular: target QED/logP/SA values are computed externally with RDKit/MoFlow code on generated molecules, so those results provide independent empirical content. The missing specification of h(z_c,z_d) and the matrix/vector dimension mismatch is a technical correctness gap, not a circularity. No load-bearing self-citation or imported uniqueness theorem was found. Because the central image-disentanglement evidence reduces partly to the supervised construction, the score is 6 rather than 0-2.

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

The central construction rests on learned Lie basis matrices {Ai} (free parameters, dimension |A|) and empirically tuned loss weights; no physically new entities (particles, forces, dimensions) are introduced. The key unsupported premise is that the matrix-exponential latent can be mapped bijectively into the flow's Euclidean latent space with the claimed factor-wise control.

free parameters (4)
  • Loss weights λGP, λP, λF, λH, λC = QM9: 20/0.1/1/10/20; ZINC: 10/0.5/1/5/20; MOSES: 10/0.5/1/5/20 (Tables 5-7)
    Empirically selected per dataset with no selection procedure described; they define the headline objective in Eq. (5).
  • Number of Lie basis T = 8 (QM9), 6 (ZINC), 6 (MOSES); 5 (dSprites) and 10 (MNIST) implied by the factor counts
    Sets the number of controllable factors; chosen per dataset without justification.
  • Lie basis matrix dimension |A| = 7×7 (QM9), 40×40 (ZINC), 29×29 (MOSES)
    Defines the dimensionality of the learned Ai and hence of the matrix exponential output; the relation to the flow latent dimension is unexplained.
  • Noise vector dimension |ε| = 10 (QM9), 20 (ZINC), 20 (MOSES)
    Input to the coupled block; hand-set.
assumptions (4)
  • domain assumption The flow latent space z is (approximately) N(0,I); the WGAN+predictor alignment is sufficient for pθ(z|c) ≈ pφ(z|c) and hence for controlled generation (Sec. 3.1, items (1)-(2))
    The implication is asserted, not proven: matching marginals and class posteriors pointwise is a heuristic goal; no convergence or error bound ties the alignment losses to faithful conditional sampling.
  • ad hoc to paper exp(Σ_i t_i A_i) with commuting Ai yields one-factor-per-coordinate control inside the FLOW's latent space after the combination block h
    Lie theory guarantees factorization of the generator's own parameterization, not that the adversarially aligned z-space coordinates are one-to-one with generative factors; DCI/FVM = 1.00 is presented as a consequence without derivation (Sec. 4.1, Eq. (5), Algorithm 1).
  • domain assumption The factor labels (one-hot dSprites factors, QED/logP/SA property values) are valid conditions for measuring disentanglement
    The metrics then measure label recoverability rather than emergent structure; comparing against unsupervised baselines (Table 1b) conflates supervised attribution with unsupervised disentanglement.
  • domain assumption Multi-scale flow splits: 'half of the pixel values are obtained from the prior' (Sec. 9), so part of z is a fixed Gaussian sample that the conditional generator must also match
    The paper does not specify how the Lie generator accounts for split-off prior factors; it notes residual shape variance from the multi-scale structure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decafs: Disentangled Conditional adversarial Flows." pith.science (2026). https://pith.science/paper/J2B75WB2

@misc{pith2026260718755,
  author       = {Pith},
  title        = {Pith review of: Decafs: Disentangled Conditional adversarial Flows},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J2B75WB2}},
  note         = {Machine review of arXiv:2607.18755}
}
read the original abstract

Flow-based models have established state-of-the-art performance in generative modeling across domains, but are hard to interpret due to their complex latent embeddings. In particular, the entanglement of generative factors in the latent space hinders controlled generation. We circumvent this issue by appealing to a novel conditional generator based on Lie groups that disentangles an alternative latent space, which is aligned closely with the latent flow space using an adversarial loss. Our approach facilitates interpretable conditional generation while obviating the need to expand the dimensionality of the flow space (owing to its invertibility requirements). The proposed model demonstrates strong performance across conditional image (including, outperforming StyleGAN on MNIST, dSprites) and molecule (using standard QM9, ZINC and MOSES) generation tasks

Figures

Figures reproduced from arXiv: 2607.18755 by the authors.

Figure 1
Figure 1. Decaf learns dis￾entangled view of flow latent space by modelling each la￾tent vector (equivalently, each point on the manifold) in terms of the coefficients along the Lie group basis {A1, A2}, representing two independent generative vectors. Since the resulting vector may be off the manifold, an exponential map projects this vector back to the nearest point on the manifold. It should be noted that the fig￾ure only … view at source ↗
Figure 2
Figure 2. Left:Decaf Architecture, comprises of four primary blocks: A Flow network which learns an invertible transformation between true data distribution and the latent space, A Lie-group generator which generates a latent space conditioned on the input labels, A discriminator Network that discriminates the flow latent space and the generated latent space for adversarial training, A predictor network that aids the flow and… view at source ↗
Figure 3
Figure 3. Left: Disentangled conditional generation for dSprites images using Decaf. Each row is obtained by varying the contribution of a specified Lie basis. Lie group coordinate t0 controls the shape between oval, heart and square, t1 controls the scale of the generated image, t2 controls the orientation, t3 controls the position on the X-axis and t4 controls the position on the Y-axis. Right: (a)-(e): MNIST images generat… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: We perform controllable genera￾tion of molecules using latent traversal in Lie group space by fixing the input noise and only varying a single property at a time. The molecule highlighted with the red bounding box is the starting point. We vary only Top row: QED Middle…
Figure 6
Figure 6. Figure 6: Coordinates tis distribution for the corresponding Lie group basis for MNIST images. We report the mean and variance of each coordinate over 10,000 noise vectors per label. The coordinates with no variance control the shape of the digit for an appropriate label and coo…
Figure 7
Figure 7. Figure 7: (a) Conditional generation of MNIST images with Lie Generator for different classes (b) Disentangled image manipulation by varying a single Lie group basis for each class. We start from the first image in each row and gradually vary a single Lie coordinate. The disenta…
Figure 8
Figure 8. Figure 8: Disentangled conditional generation with other shapes as default vector. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Sample density from (a) dSprites dataset (b) prior of normalizing flow (c) randomly [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Caption 17 [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Conditional generation with QM9 The joint distribution over molecular graphs p(A, B) can be decomposed as p(A, B) = p(A|B)p(B). Both distributions are modelled by independent flows, log p(A, B) = log p(A|B) + log p(B) (10) = log N (fA|B(A, B)|0, I) +X TA t=1 log [PIT…
Figure 12
Figure 12. Figure 12: Conditional generation with Moses dataset. (a)-(c) Unconditional generation. (d)-(f) [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]
Figure 13
Figure 13. Figure 13: Molecules generated by unconstrained penalized logP generation from decaf. [PITH_FULL_IMAGE:figures/full_fig_p024_13.png]
Figure 14
Figure 14. Figure 14: Molecules generated by unconstrained QED generation from decaf. [PITH_FULL_IMAGE:figures/full_fig_p025_14.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 1 canonical work pages

  1. [1]

    Deep variational information bottleneck.arXiv preprint arXiv:1612.00410,

    Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck.arXiv preprint arXiv:1612.00410,

    arXiv

  2. [7]

    RDKit: Open-source cheminformatics

    Gregory Landrum. RDKit: Open-source cheminformatics. Release 2014.03.1, June

    2014

  3. [10]

    A property-guided diffusion model for generating molecular graphs

    Changsheng Ma, Taicheng Guo, Qiang Yang, Xiuying Chen, Xin Gao, Shangsong Liang, Nitesh Chawla, and Xiangliang Zhang. A property-guided diffusion model for generating molecular graphs. InICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 2365–2369,

    2024

  4. [13]

    Trung Ngo, Najwa Laabid, Ville Hautamäki, and Merja Heinäniemi

    URL https://openreview.net/forum?id= BkxthxHYvr. Trung Ngo, Najwa Laabid, Ville Hautamäki, and Merja Heinäniemi. The transitive information theory and its application to deep generative models.arXiv preprint arXiv:2203.05074,

    arXiv

  5. [14]

    Improving denoising diffusion probabilistic models via exploiting shared representations.arXiv preprint arXiv:2311.16353,

    Delaram Pirhayatifard, Mohammad Taha Toghani, Guha Balakrishnan, and César A Uribe. Improving denoising diffusion probabilistic models via exploiting shared representations.arXiv preprint arXiv:2311.16353,

    arXiv

  6. [15]

    Graphaf: a flow-based autoregressive model for molecular graph generation.arXiv preprint arXiv:2001.09382,

    Chence Shi, Minkai Xu, Zhaocheng Zhu, Weinan Zhang, Ming Zhang, and Jian Tang. Graphaf: a flow-based autoregressive model for molecular graph generation.arXiv preprint arXiv:2001.09382,

    arXiv 2001

  7. [16]

    Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456,

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456,

    arXiv 2011

  8. [17]

    Abode: Ab initio antibody design using conjoined odes.arXiv preprint arXiv:2306.01005,

    Yogesh Verma, Markus Heinonen, and Vikas Garg. Abode: Ab initio antibody design using conjoined odes.arXiv preprint arXiv:2306.01005,

    arXiv

Show all 22 references
  1. [19]

    Uncovering the disentanglement capability in text-to-image diffusion models

    Qiucheng Wu, Yujian Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. ...

    1900

  2. [21]

    Graph convolutional policy network for goal-directed molecular graph generation.arXiv preprint arXiv:1806.02473,

    Jiaxuan You, Bowen Liu, Rex Ying, Vijay Pande, and Jure Leskovec. Graph convolutional policy network for goal-directed molecular graph generation.arXiv preprint arXiv:1806.02473,

    arXiv

  3. [32]

    We train both flows using Adam optimizer with learning rate1e−3for 100 epochs. 10.1 DISENTANGLEMENT MATRICES We rely on various popular metrics such as FactorV AE Metric(FVM) Kim & Mnih (2018), DCI Eastwood & Williams (2018), Seperated Attribute Predictability (SAP) score Kuma...

    2018

  4. [2013]

    Variational inference of disentangled latent concepts from unlabeled observations.arXiv preprint arXiv:1711.00848,

    Abhishek Kumar, Prasanna Sattigeri, and Avinash Balakrishnan. Variational inference of disentangled latent concepts from unlabeled observations.arXiv preprint arXiv:1711.00848,

    arXiv

  5. [2014]

    Density estimation using real nvp.arXiv preprint arXiv:1605.08803,

    Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp.arXiv preprint arXiv:1605.08803,

    arXiv

  6. [2015]

    Yann LeCun

    URL https://doi.org/10.5281/zenodo.10398. Yann LeCun. The mnist database of handwritten digits.http://yann. lecun. com/exdb/mnist/,

    doi

  7. [2017]

    Jonathan Ho and Tim Salimans

    URLhttps://openreview.net/forum?id=Sy2fzU9gl. Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598,

    arXiv

  8. [2018]

    Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

    arXiv

  9. [2019]

    Disdiff: Unsupervised disentanglement of diffusion probabilistic models.arXiv preprint arXiv:2301.13721,

    Tao Yang, Yuwang Wang, Yan Lv, and Nanning Zheng. Disdiff: Unsupervised disentanglement of diffusion probabilistic models.arXiv preprint arXiv:2301.13721,

    arXiv

  10. [2020]

    Categorical normalizing flows via continuous transformations

    Phillip Lippe and Efstratios Gavves. Categorical normalizing flows via continuous transformations. arXiv preprint arXiv:2006.09790,

    2006 arXiv

  11. [2021]

    Nice: Non-linear independent components estimation.arXiv preprint arXiv:1410.8516,

    Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation.arXiv preprint arXiv:1410.8516,

    arXiv

  12. [2022]

    Conditional generative adversarial nets.arXiv preprint arXiv:1411.1784,

    Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets.arXiv preprint arXiv:1411.1784,

    arXiv

  13. [2023]

    Improving robustness and generality of nlp models using disentangled representations.arXiv preprint arXiv:2009.09587,

    Jiawei Wu, Xiaoya Li, Xiang Ao, Yuxian Meng, Fei Wu, and Jiwei Li. Improving robustness and generality of nlp models using disentangled representations.arXiv preprint arXiv:2009.09587,

    2009 arXiv

  14. [2024]

    Emile Mathieu, Tom Rainforth, N Siddharth, and Yee Whye Teh

    doi: 10.1109/ICASSP48485.2024.10447350. Emile Mathieu, Tom Rainforth, N Siddharth, and Yee Whye Teh. Disentangling disentanglement in variational autoencoders. InInternational Conference on Machine Learning, pp. 4402–4412,

    2024

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.