Pith. sign in

REVIEW 5 major objections 5 minor 12 references

CoQui: A Coordinate-Conditioned Quantum Implicit Generative Adversarial Network for End-to-End Image Generation

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

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

Pith's one-line read CoQui reformulates quantum image generation as coordinate-conditioned implicit function learning, using a five-qubit circuit that reads each pixel from a color qubit and beats amplitude-mapping QGAN baselines in simulation.

desk verdict Clean architectural idea for quantum image generation with a real resource argument, but the abstract's empirical claims outrun the paper's own tables. read the letter →

arxiv 2608.11884 v1 pith:65MH6LTA submitted 2026-08-12 quant-ph cs.AIcs.CV

Xue Yang , Rigui Zhou , ShiZheng Jia , Dax Enshan Koh , Siong Thye Goh , Young-Wook Cho , YaoChong Li , Xuezhi Ma
show 2 more authors
Hongyu Chen Xin Wang
This is my paper · ORCID
classification quant-phcs.AIcs.CV
keywords quantumgenerativeadversarialnetworksimplicitneuralrepresentationcoordinate-conditionedimagegenerationdatare-uploadingvariationalcircuitpixelreadoutWassersteinGAN
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

CoQui argues that quantum image generation should stop assigning pixel intensities to normalized quantum-state amplitudes and instead treat an image as a function from pixel coordinates to intensities. The generator takes a coordinate and a shared latent code, turns them into circuit angles through a classical embedding network, evaluates a small variational circuit, and reads the pixel value from the Pauli-$Z$ expectation of a dedicated color qubit; an image is assembled by querying every coordinate. In simulated experiments on MNIST and Fashion-MNIST this five-qubit design beats two amplitude-mapping QGAN baselines in FID and related metrics while using fewer qubits, and it matches or beats a classical INR-GAN of similar parameter scale. The stakes are that quantum image generation can be decoupled from image resolution, with the quantum circuit responsible for actual pixel-level modeling rather than just a compressed latent.

What carries the argument

The central object is the quantum implicit generator $G_\Phi(c,z)$, a variational circuit with one color qubit $q_0$ plus $N_f$ feature qubits, evaluated once per pixel coordinate. Its structure layers a brightness-bias rotation on the color qubit, scaled data re-uploading (per-layer trainable scale $s^l$ and bias $b^l$ applied to the coordinate-and-latent angles), local rotations, ring-CNOT entanglement among feature qubits, feature-to-color writing via controlled-$R_Y$ gates from each feature qubit to the color qubit, and residual rotations on the color qubit; the output is the Pauli-$Z$ expectation of $q_0$. This circuit carries the argument because it turns each pixel into an explicit modular function of location and latent code, so resolution growth is absorbed by additional coordinate queries rather than by additional qubits or by redistributing a fixed probability mass.

What would settle it

Train the exact five-qubit, twenty-layer CoQui generator alone, without the GAN, under pixel-wise supervision on a 64×64 natural image or a high-frequency synthetic image; if its reconstruction error is substantially worse than a classical implicit network with similar parameters, the coordinate-query scheme cannot claim to hold beyond small grayscale benchmarks, and most of the demonstrated quality must come from the classical components.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that replacing the amplitude-mapping readout with a coordinate-conditioned implicit readout eliminates the two structural defects of earlier QGAN image generators. The generator $G_\Phi(c,z)$ is defined by a five-qubit circuit (one color qubit, four feature qubits, twenty layers): for each normalized coordinate $c=(x,y)$ and a shared latent $z$, a classical embedding network produces angles $A(c,z)$, layer-wise scales and biases turn these into injected rotations, feature-qubit controlled-$R_Y$ gates write feature information onto the color qubit, and the pixel intensity is $G_\Phi(c,z)=(1-\langle Z_0\rangle)/2$. Because each pixel comes from an independent circuit query, there is no shared probability-normalization constraint across pixels; because the circuit has no address qubits, resolution is decoupled from qubit count. The paper reports that this design outperforms FRQI-style and PQWGAN generation in visual and quantitative quality at 28×28 in simulation while using five qubits and one circuit, and that it also beats the corresponding classical baseline.

Load-bearing premise

The load-bearing premise is that a five-qubit, twenty-layer coordinate-conditioned circuit is expressive and trainable enough to reproduce full-image statistics, so the quantum circuit—not the classical embedding network or the classical critic—is doing the generative work; the paper gives no expressivity or trainability guarantee, and its own capacity ablation shows that slightly changing the circuit configuration can substantially worsen FID.

Editorial extensions

If this is right

  • Image resolution no longer drives qubit count: the same generator can in principle render any grid by more coordinate queries.
  • Amplitude-mapping normalization artifacts, such as the lower overall brightness observed in PQWGAN, are structurally avoided because each pixel is read from its own expectation value.
  • The feature-to-color writing ansatz becomes a candidate building block for other coordinate-conditioned quantum models, including super-resolution and continuous signal generation.
  • The comparison against a classical INR-GAN at comparable parameter scale gives a concrete benchmark for judging whether a quantum generator's expressivity earns its cost.

Reading between the lines

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

  • The paper does not count circuit executions: one 28×28 image requires 784 forward passes through a 20-layer circuit, so on real hardware the practical cost shifts from qubit count to per-pixel query time and finite-shot noise.
  • The paper's own conclusion flags that all results are classical simulation on grayscale benchmarks; a fair hardware comparison must include measurement-shot noise and device connectivity, which the reported FID numbers do not cover.
  • A clean test of where the expressivity lives would be to train CoQui's classical embedding plus a fixed random quantum circuit, or a classical MLP with the same coordinate and latent conditioning, and compare FID: if either matches CoQui, the quantum generator's contribution to image quality is smaller than the paper asserts.
  • Because the readout is a single scalar per coordinate, the same recipe extends naturally to audio, point clouds, and implicit surfaces, where a continuous coordinate-to-signal map is wanted.
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 CoQui, a quantum generative adversarial network for image generation in which a variational quantum circuit is queried at each spatial coordinate, conditioned on a latent code shared across the image, and pixel intensity is read from the expectation value of a dedicated color qubit. The circuit uses one color qubit plus N_f feature qubits with data re-uploading, so the number of qubits does not depend on image resolution. The authors report simulated experiments on MNIST and Fashion-MNIST comparing against PQWGAN, the Wasserstein QGAN of Jäger et al., and a classical INR-GAN baseline, together with ablations of circuit components and capacity. The central claims are that the coordinate-query design decouples qubit count from resolution, avoids inter-pixel probability competition, and achieves better generation quality than the quantum baselines and, per the abstract, than the corresponding classical baseline.

Significance. If the results are reproducible, the coordinate-conditioned readout scheme is a clean architectural contribution: it removes the resolution dependence of the qubit count and replaces joint amplitude decoding with independent pixel queries, which is a principled way to sidestep normalization coupling. The paper provides detailed ablations (Tables 3–5) and a resource comparison (Table 2) that support the feasibility of a five-qubit quantum generator. However, the claimed advantage over the classical INR-GAN baseline is not supported by the reported numbers; at best CoQui is competitive. The structural insight is useful, and the comparison against amplitude-mapping QGANs is favorable, but the paper's headline empirical claim must be revised to match its own evidence.

major comments (5)
  1. [Abstract; Main Results, Figure 2, Table 1] The statement that CoQui 'achieves better generation quality than the corresponding classical baseline' is contradicted by the paper's own results. On full-dataset MNIST, Classical INR-GAN obtains FID 34 versus CoQui FID 42 (Figure 2); on Fashion-MNIST the two are tied at 72. In the class-wise MNIST table, Classical INR-GAN has a lower FID on six of the ten classes (1, 2, 3, 5, 6, 7), and the paper's own text concedes CoQui is best or tied-best on only four MNIST classes. The conclusion itself downgrades the claim to 'competitive generative performance.' The abstract, contributions, and any summary statements should be reworded to match the evidence.
  2. [Equation (3); Appendix 1] The manuscript gives three different descriptions of the angle output transformation. Equation (3) applies α tanh; Appendix 1 says the raw angles are 'squashed by atanh nonlinearity and scaled by π'; and Table A1 lists 'π tanh(·)'. The value of α is never specified. Please state the exact nonlinearity, scaling factor, and parameter value used to produce all reported results; as written, the evaluated configuration is ambiguous and not reproducible.
  3. [Tables 3, 4, and 5; Figure 2] The reported FID values for the same nominal configuration are inconsistent. Tables 3 and 4 give Full CoQui FID 40.15 ± 3.45, while Table 5 reports FID 42.59 for N_f = 4, r = 20, and Figure 2 reports FID 42 for full-dataset MNIST. If these numbers correspond to different evaluation settings (class-wise versus multi-class, different seeds, or different data subsets), this must be stated explicitly; otherwise the ablation results are internally contradictory and the main comparisons cannot be verified.
  4. [Table 1 and Figure 2] The headline FID values in Table 1 and Figure 2 are reported without error bars or a number of seeds, while the ablation tables report mean ± standard deviation. Given the small training set (1000 images) and the known variance of FID estimates, please report means and standard deviations over at least three independent runs for all central numbers, including the class-wise and full-dataset comparisons.
  5. [Experiments; Classical INR-GAN baseline] The 'Classical INR-GAN' baseline is described only as a 'classical counterpart of CoQui' and no architecture, parameter count, or training details are given. Because the quantum-versus-classical comparison and the phrase 'comparable parameters scale' rest on this baseline, please provide a full description of the classical model and its parameter count, or point to a specific implementation and configuration.
minor comments (5)
  1. [Table 1 caption] The caption states 'Best results are in bold,' but no entries in the typeset table are bold.
  2. [Experiments, Setup] The sentence 'The training dataset consists of 1000 real images in each experiment' is ambiguous: please clarify whether the 10-class MNIST run uses 1000 images total or 1000 per class.
  3. [Method and Experiments] The circuit depth is denoted L in the method section and r in the experimental setup; please use one symbol consistently.
  4. [Table 5, capacity ablation] The ablation shows non-monotonic FID as a function of N_f (N_f=2: 39.53, N_f=4: 42.59, N_f=5: 37.88, N_f=6: 44.65), so the chosen N_f=4 is not the optimum and the configuration appears fragile; a sentence acknowledging this and discussing optimization difficulty would help.
  5. [Figure 2] The figure is dense and the per-method FID labels are easily confused; consider separating the panels or moving the numbers into a table with confidence intervals.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the quantum generator's output is defined by explicit circuit equations and evaluated on held-out samples; the only data-derived initialization is a design convention, not a prediction.

full rationale

The paper's derivation chain is empirical and self-contained rather than circular. The quantum generator is explicitly constructed through circuit definitions (Eqs. 6-20), with pixel intensities defined as G(c,z) = (1 - <Z0>)/2, and the reported FID/JSD/Entropy metrics are measured on generated samples against the real data distribution. The brightness-bias initialization beta0 = arccos(1 - 2 mu0) is intentionally chosen so that the initial readout equals a preset mean pixel value mu0; this is an initialization convention, not a fitted parameter that is subsequently relabeled as a prediction. Hyperparameter selection in Table 5 is standard model selection based on validation metrics, and the final configuration is then evaluated independently. The self-citations (Yang et al. 2026a, 2026b) appear only in Related Work and are contextual; no load-bearing argument relies on them, and no uniqueness theorem is imported from the authors' prior work. The abstract's claim that CoQui 'achieves better generation quality than the corresponding classical baseline' is not supported by the paper's own tables (full-dataset MNIST FID 42 vs 34; class-wise Classical INR-GAN wins on six of ten MNIST classes), and the conclusion downgrades this to 'competitive.' In addition, Eq. (3) uses tanh with scaling alpha while Appendix 1 says atanh and pi for the same output transformation, creating ambiguity about the exact evaluated configuration. These are internal-consistency and reproducibility concerns, not circularity. No step in the paper reduces, by construction or self-citation, to its own inputs.

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

This ledger lists the implementation choices and background assumptions the reported results depend on. The main hand-selected constants are the circuit size (Nf=4), depth (L=20), positional encoding bands (K=6), latent dimension (10), and the brightness prior mu0 fitted to dataset mean. These values were chosen through ablations on the test benchmarks rather than derived from a theory. The core scientific assumptions are that exact statevector simulation is a faithful proxy for the device, that the re-uploading PQC is expressive enough for the image function, and that WGAN-GP training of the hybrid model converges.

free parameters (6)
  • Feature qubit count Nf = 4
    Selected via capacity ablation on MNIST (Table 5); Nf=5 gives lower FID but Nf=4 chosen for best class JSD/Entropy.
  • Circuit depth L = 20
    Chosen from ablation (Table 5); 10-layer variant has similar or better FID (39.07) but 20 used in final model.
  • Positional encoding frequency bands K = 6
    Appendix 1 states six frequency bands yielding 26-dim coordinate feature; chosen by hand.
  • Latent dimension dz = 10
    Appendix 1; chosen by hand.
  • Angle scaling factor alpha = unspecified; Appendix says pi
    Eq. (3) uses alpha but value not given in main text; Appendix 1 says scaled by pi.
  • Brightness initialization mean mu0 = preset mean pixel value
    Set to the mean pixel intensity of the training data (Eq. 7); this is a data-fitted constant.
assumptions (4)
  • domain assumption Statevector simulation of the PQC is exact and noise-free
    All experiments run in classical simulation; hardware noise and finite-shot effects are deferred (Conclusion).
  • domain assumption A five-qubit data-re-uploading PQC can represent the target image functions
    The generator maps coordinates+latent to pixel intensities via the PQC (Eqs. 18-20); no expressivity bound is given, and Table 5 shows sensitivity to Nf.
  • domain assumption WGAN-GP training on a hybrid quantum-classical generator converges to a useful optimum
    The paper relies on standard GAN training stability assumptions; no convergence guarantee for the PQC parameter updates is provided.
  • standard math The classical embedding network and PQC are jointly trainable by backpropagation through exact expectation values
    In simulation, gradients of <Z0> with respect to circuit parameters are well-defined via parameter-shift rules; the paper does not state the differentiation method.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CoQui: A Coordinate-Conditioned Quantum Implicit Generative Adversarial Network for End-to-End Image Generation." pith.science (2026). https://pith.science/paper/65MH6LTA

@misc{pith2026260811884,
  author       = {Pith},
  title        = {Pith review of: CoQui: A Coordinate-Conditioned Quantum Implicit Generative Adversarial Network for End-to-End Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/65MH6LTA}},
  note         = {Machine review of arXiv:2608.11884}
}
read the original abstract

Quantum generative adversarial networks (QGANs) have attracted increasing attention for image generation using parameterized quantum circuits. Existing amplitude-based approaches face two key limitations: pixel locations are typically encoded by computational-basis indices or address qubits, causing quantum resources to grow with image resolution; meanwhile, jointly decoding many pixels from normalized quantum states introduces probability competition among pixels and limits precise pixel-wise control. To address these issues, we reformulate quantum image generation as coordinate-conditioned implicit function learning. Our method takes spatial coordinates and latent variables as inputs, uses a classical embedding network to generate input-dependent circuit parameters, and evaluates a variational quantum circuit at each coordinate. Pixel intensities are directly obtained from the expectation value of a dedicated color qubit, and a complete image is generated by querying all spatial coordinates. This design decouples image resolution from address-qubit requirements and avoids shared probability-normalization constraints across pixels. We further design a specialized variational quantum circuit to provide structural inductive bias for coordinate-conditioned generation. Simulated experiments on two benchmark datasets show that our method outperforms FRQI-based generation and PQWGAN in visual and quantitative quality while using fewer qubits, and also achieves better generation quality than the corresponding classical baseline.

Figures

Figures reproduced from arXiv: 2608.11884 by the authors.

Figure 1
Figure 1. Overview of the CoQui framework. (a) Overall architecture: Given a spatial coordinate c = (x, y) and a latent code z, CoQui utilizes a classical embedding network Γ to generate coordinate- and latent-conditioned parameters A(c, z) ∈ R Nf ×3 . These are transformed into layer-wise scaled parameters Θi,ℓ for the quantum generator G. The pixel intensity is computed as G(c, z) = 1−⟨Z0⟩ 2 by measuring the Pauli-Z expecta… view at source ↗
Figure 2
Figure 2. Qualitative comparison on MNIST and Fashion-MNIST. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Representative MNIST samples for (a) CoQui and its [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Representative MNIST samples for two capacity studies. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 1
Figure 1. Figure 1: Training FID curves of the jointly trained MNIST 0–9 model and the individual digit-specific models. [PITH_FULL_IMAGE:figures/full_fig_p010_1.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 10 canonical work pages

  1. [1]

    Wasserstein Generative Adversarial Networks

    Arjovsky,M.;Chintala,S.;andBottou,L.2017. Wasserstein Generative Adversarial Networks. In Precup, D.; and Teh, Y.W.,eds.,Proceedingsofthe34thInternationalConference onMachineLearning,volume70ofProceedingsofMachine Learning Research, 214–223. PMLR. Cao, J.; Wang, Q.; Xian, Y.; Li, Y.; Ni, B.; Pi, Z.; Zhang, K.; Zhang, Y.; Timofte, R.; and Van Gool, L

    work page 2017

  2. [4]

    InICASSP 2023-2023 IEEEinternationalconferenceonacoustics,speechandsig- nal processing (ICASSP), 1–5

    IQ- GAN: Robust quantum generative adversarial network for image synthesis on NISQ devices. InICASSP 2023-2023 IEEEinternationalconferenceonacoustics,speechandsig- nal processing (ICASSP), 1–5. IEEE. Dallaire-Demers, P.-L.; and Killoran, N

    work page 2023

  3. [8]

    Mescheder,L.;Oechsle,M.;Niemeyer,M.;Nowozin,S.;and Geiger,A.2019

    Quantum generative adversariallearning.Physicalreviewletters,121(4):040502. Mescheder,L.;Oechsle,M.;Niemeyer,M.;Nowozin,S.;and Geiger,A.2019. OccupancyNetworks:Learning3DRecon- structioninFunctionSpace.InProceedingsoftheIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Müller, T.; Evans, A.; Schied, C.; and Keller, A

    work page 2019

  4. [12]

    InProceedings of the 41st International Conference on Machine Learning, ICML’24

    Quantum implicit neural representations. InProceedings of the 41st International Conference on Machine Learning, ICML’24. JMLR.org. Zoufal,C.;Lucchi,A.;andWoerner,S.2019. Quantumgen- erativeadversarialnetworksforlearningandloadingrandom distributions.npj Quantum Information, 5(1):

    work page 2019

  5. [69]

    ISBN 978-3-032- 15931-1

    Cham: Springer Nature Switzerland. ISBN 978-3-032- 15931-1. Gonzalez, R. C. 2009.Digital image processing. Pearson education india. Goodfellow,I.;Pouget-Abadie,J.;Mirza,M.;Xu,B.;Warde- Farley, D.; Ozair, S.; Courville, A.; and Bengio, Y

    work page 2009

  6. [91]

    Hybridquantum–classicalgenerativeadversarialnetworkfor high-resolution image generation.IEEE Transactions on Quantum Engineering, 4: 1–19

    Tsang,S.L.;West,M.T.;Erfani,S.M.;andUsman,M.2023. Hybridquantum–classicalgenerativeadversarialnetworkfor high-resolution image generation.IEEE Transactions on Quantum Engineering, 4: 1–19. Vieloszynski, A.; Cherkaoui, S.; Ahmad, O.; Laprade, J.- F.; Nahman-Lévesque, O.; Aaraba, A.; and Wang, S

    work page 2023

  7. [226]

    ImprovedTechniquesforTrain- ing GANs

    Salimans,T.;Goodfellow,I.;Zaremba,W.;Cheung,V.;Rad- ford,A.;andChen,X.2016. ImprovedTechniquesforTrain- ing GANs. InAdvances in Neural Information Processing Systems, volume

    work page 2016

  8. [2016]

    Chen, Y.; Liu, S.; and Wang, X

    Mode Regularized Generative Adversarial Networks.arXiv preprint arXiv:1612.02136. Chen, Y.; Liu, S.; and Wang, X

    arXiv

Show all 12 references
  1. [2018]

    Quantum generative adversarial networks.Phys. Rev. A, 98: 012324. Eren,S.M.2026. ImplementationofQuantumImplicitNeu- ralRepresentationinDeterministicandProbabilisticAutoen- coders for Image Reconstruction/Generation Tasks.arXiv preprint arXiv:2603.06755. Fujihashi, T.; and Koi...

    2026

  2. [2023]

    In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1796–1807

    CiaoSR: Continuous Implicit Attention-in-Attention Net- work for Arbitrary-Scale Image Super-Resolution. In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1796–1807. Chang, S. Y.; Thanasilp, S.; Saux, B. L.; Vallecorsa, S.; and Grossi,M.2024. Latent...

    2024 arXiv

  3. [2024]

    In2024 IEEE 10th World Forum on Internet of Things (WF-IoT), 1–7

    Latentqgan: A hybrid qgan with classical convolutional au- toencoder. In2024 IEEE 10th World Forum on Internet of Things (WF-IoT), 1–7. IEEE. Wang,S.;Theobalt,C.;andGolyanik,V.2026. Quantumvi- sualfieldswithneuralamplitudeencoding.AdvancesinNeu- ral Information Processing Syst...

    2026

  4. [2025]

    Jäger, J.; Kiwit, F

    QFGN: A Quantum Approach to High-Fidelity Implicit Neural Repre- sentations.arXiv preprint arXiv:2504.19053. Jäger, J.; Kiwit, F. J.; and Riofrío, C. A

    arXiv

Pith tools

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