{"id":"a5c03c2a-dbdf-4e8c-ba09-78070e2caba4","arxiv_id":"2608.11884","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CoQui is a coordinate-conditioned quantum implicit GAN that reads each pixel from a dedicated color qubit's expectation value, decoupling qubit count from image resolution and reporting improved generation over amplitude-based quantum GAN baselines.","lead":"CoQui generates images by asking a tiny five-qubit quantum circuit to output one pixel at a time, conditioning on the pixel's coordinates and a shared latent code. The paper claims this avoids the pixel-coupling problem of amplitude-based quantum GANs and uses fewer qubits, while showing mixed quality versus a classical equivalent.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's claim that CoQui outperforms the classical baseline contradicts the paper's own FID tables (full-dataset MNIST 42 vs 34; class-wise classical wins 6/10), so the central empirical superiority claim is not supported.","rationale":"I read the paper as making two separable claims: (1) coordinate-conditioned per-pixel readout decouples qubit count from resolution and removes shared normalization; (2) in simulation this yields better generation quality than amplitude-mapping QGANs and, per the abstract, than the classical INR baseline. Claim (1) follows from the construction and is not disputed. Claim (2) is the empirical payoff, and the paper's own numbers contradict the classical-baseline part: the full-dataset MNIST FID is worse by 8 points, class-wise results favor the classical model on most MNIST classes, and Fashion-MNIST is tied. The reader's designated weakest assumption is circuit expressivity/trainability; that is a legitimate open question, but the more immediate, load-bearing weakness is the internal mismatch between the abstract and the evidence. This is a correctness risk of the advertised contribution, not a disagreement with consensus. A conditional acceptance requiring corrected claims and seed-averaged, code-checkable comparisons is the appropriate outcome, matching the reader's verdict; I would not move it to reject because the architectural idea remains coherent and the qualitative ablations are informative. The Appendix/main-text angle-mapping inconsistency should be fixed in the same revision.","tokens_in":13605,"tokens_out":9015,"duration_ms":92910,"concrete_test":"Run the full 10-class MNIST experiment under the exact reported protocol (1000 training images, 1000 epochs, batch size 5, same FID reference set) for at least five independently seeded trials and report mean ± std FID for CoQui and the Classical INR-GAN, plus the class-wise counts of which method has lower FID per class. If the classical baseline's mean FID remains lower than CoQui's, the abstract and contribution statements must be revised from 'better/outperforms' to 'comparable/competitive' before the empirical claims can be accepted as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is not only architectural; the abstract and contributions assert that CoQui 'achieves better generation quality than the corresponding classical baseline,' while the conclusion downgrades this to 'competitive.' The reported evidence does not support the stronger assertion. In the full-dataset setting, Figure 2 shows Classical INR-GAN FID = 34 on MNIST versus CoQui FID = 42, and on Fashion-MNIST the two are tied at 72. In Table 1's class-wise MNIST results, Classical INR-GAN has lower FID on six of ten classes (1, 2, 3, 5, 6, 7), one tie (8), and CoQui is better on three (0, 4, 9); the authors' own text concedes CoQui is best or tied-best on only four MNIST classes. This is an internal inconsistency in the evidence that directly undercuts the stated empirical advantage over the classical counterpart. It is compounded by reproducibility problems: Eq. (3) uses tanh with scaling factor alpha, while Appendix 1 says 'atanh' and 'pi' for the same output transformation, so the exact evaluated configuration is ambiguous. With no error bars and no released code, the reported FID differences cannot be checked.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":13825,"tokens_out":6781,"duration_ms":62771,"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":[{"comment":"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.","section":"Abstract; Main Results, Figure 2, Table 1"},{"comment":"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.","section":"Equation (3); Appendix 1"},{"comment":"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.","section":"Tables 3, 4, and 5; Figure 2"},{"comment":"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.","section":"Table 1 and Figure 2"},{"comment":"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.","section":"Experiments; Classical INR-GAN baseline"}],"minor_comments":[{"comment":"The caption states 'Best results are in bold,' but no entries in the typeset table are bold.","section":"Table 1 caption"},{"comment":"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.","section":"Experiments, Setup"},{"comment":"The circuit depth is denoted L in the method section and r in the experimental setup; please use one symbol consistently.","section":"Method and Experiments"},{"comment":"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.","section":"Table 5, capacity ablation"},{"comment":"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.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The abstract and contributions overstate the empirical advantage over the classical baseline; this is an internal inconsistency rather than a matter of interpretation. The authors should be asked to reconcile the abstract with the reported FID values. Also, the code is not yet public; given the small scale of the experiments, releasing code and hyperparameters (including the angle scaling factor α) is important for verification. The architectural contribution is sound and the comparison with quantum baselines is favorable, so the paper is worth revising carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main thing to know: this paper has a genuinely sensible architecture, and the decoupling claim holds. Evaluating a small variational circuit per pixel with a dedicated color qubit readout does fix the qubit count independent of resolution and removes the shared-normalization coupling that hampers amplitude-based methods. That is a real contribution. The circuit design with feature-to-color controlled-RY writing, scaled data re-uploading, and a brightness-bias initialization is well motivated, and the ablations show the components matter. On the resource side, five qubits versus 192 for PQWGAN is a strong selling point. The soft spots are mostly about claims versus evidence. The abstract says CoQui achieves better generation quality than the corresponding classical baseline, but the full-dataset MNIST FID is 42 versus 34 for Classical INR-GAN, and the class-wise table has the classical baseline winning six of ten classes. The conclusion downgrades to 'competitive', but the abstract and contributions list do not. That is an internal inconsistency that has to be fixed before the empirical claims can be taken at face value. The paper would be acceptable if it claimed parity with the classical baseline while beating the quantum amplitude-mapping baselines; that is already a useful result. There are also two reproducibility issues. The main text Eq. (3) says tanh with a scaling factor alpha; Appendix 1 says atanh scaled by pi for the same output transformation. That is a concrete ambiguity in the evaluated configuration. And the main results have no error bars, even though the ablation study reports them; with no code released, I cannot check whether the FID differences are noise. The capacity ablation also shows fragility (Nf=6 makes FID worse), so I would not lean on any strong scaling claim beyond the fixed qubit count. None of this undermines the core architectural proposal. The paper is worth reading for the design and the honest ablation study. But the abstract overstates the empirical advantage, and the missing implementation details need correction. Recommendation: send it to peer review. The architectural contribution and the resource-efficiency argument deserve referee time. The authors should be required to align the abstract with the actual FID tables, clarify the angle-mapping discrepancy, and ideally release code or at least add error bars. With those fixes, this would be a solid contribution to the quantum generative modeling literature.","headline":"Clean architectural idea for quantum image generation with a real resource argument, but the abstract's empirical claims outrun the paper's own tables.","tokens_in":742,"tokens_out":645,"would_cite":true,"duration_ms":22170,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["quantum generative adversarial networks","implicit neural representation","coordinate-conditioned image generation","data re-uploading","variational quantum circuit","pixel readout","Wasserstein GAN","image generation"],"falsifier":"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.","tokens_in":13360,"feed_emoji":"⚛️","tokens_out":10248,"duration_ms":103403,"temperature":0.7,"pith_summary":"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.","feed_headline":"A five-qubit pixel-query circuit beats larger quantum image GANs","feed_subtitle":"Reading each pixel from one color qubit lowers qubit count and beats amplitude-mapping QGAN baselines.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Provides the end-to-end amplitude-mapping Wasserstein QGAN baseline that CoQui compares against under the same evaluation protocol.","marker":"(Jäger, Kiwit, and Riofrío 2026)"},{"why":"Supplies the PQWGAN patch-based amplitude baseline and the Wasserstein objective setting that CoQui inherits.","marker":"(Tsang et al. 2023)"},{"why":"Introduces quantum implicit neural representations and the coordinate-query idea that CoQui turns into a generative model.","marker":"(Zhao et al. 2024)"},{"why":"Establishes data re-uploading as a mechanism for injecting coordinate and latent information repeatedly into the circuit.","marker":"(Pérez-Salinas et al. 2020)"},{"why":"Provides the WGAN-GP gradient-penalty objective used to train CoQui's critic and generator.","marker":"(Gulrajani et al. 2017)"}],"fun_headline_variants":["Five-qubit quantum GAN sheds address qubits, beats bigger models","Implicit quantum GAN queries color qubit per pixel, outperforms","One color qubit per pixel: quantum GAN scales past resolution limits","CoQui: coordinate-conditioned QGAN cuts qubits, boosts image quality","Pixel-query quantum GAN: no address bits, better than FRQI and PQWGAN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Five-qubit quantum GAN sheds address qubits, beats bigger models","Implicit quantum GAN queries color qubit per pixel, outperforms","One color qubit per pixel: quantum GAN scales past resolution limits","CoQui: coordinate-conditioned QGAN cuts qubits, boosts image quality","Pixel-query quantum GAN: no address bits, better than FRQI and PQWGAN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000705,"raw_usage":{"total_tokens":3209,"prompt_tokens":1003,"completion_tokens":2206,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":2103}},"tokens_in":619,"tokens_out":2206,"duration_ms":16821,"temperature":1.0,"reasoning_tokens":2103,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:23:49.884676+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Hybridquantum–classicalgenerativeadversarialnetworkfor high-resolution image generation.IEEE Transactions on Quantum Engineering, 4: 1–19","cited_arxiv_id":null,"evidence_quote":"Supplies the PQWGAN patch-based amplitude baseline and the Wasserstein objective setting that CoQui inherits."}],"review_version":1}