{"id":"247ef903-c6fc-4dd5-b4c2-acc6612fd955","arxiv_id":"2411.13920","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"iHQGAN is the first quantum generative adversarial network for unsupervised image-to-image translation, using shared inverse quantum circuits and classical helper networks, with experiments on MNIST-derived edge detection, font transfer, and denoising tasks.","lead":"iHQGAN, a hybrid quantum-classical model, performs unsupervised image-to-image translation by pairing two parameter-shared quantum generators with small classical helper networks. It is the first such quantum model in this task, but the paper's training equations are internally inconsistent and the reported metrics rest on post-hoc choices.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's cycle-consistency targets are swapped relative to Eqs. (4)–(5), so the objective actually optimized is unspecified; with no code link, the reported comparisons do not support the central claim.","rationale":"The strongest claim depends on iHQGAN actually optimizing the cycle-consistency objective described in Section 3.5. The reader's weakest assumption identifies the same point: Algorithm 1 lines 22 and 28 swap the reconstruction targets compared with Eqs. (4) and (5). I consider this the single most load-bearing concern because it makes the trained objective undefined in the only detailed implementation description, and no code link is provided to disambiguate. If the printed algorithm is what was run, then the model did not enforce the claimed content-consistency constraint, so the qualitative and quantitative results do not support the paper's conclusions. If the equations are what was run, then the pseudocode is misleading enough that the reproducibility of the comparison is compromised. The other concerns noted by the reader, such as the WGAN-GP sign error, hyperparameter selection on the test set, post-processing asymmetry, and parameter-count omissions, are serious but secondary; fixing them would not suffice if the actual loss is unknown. The proposed test, checking the source code or re-running the experiment with the corrected objective, would settle whether this concern lands. Since this concern aligns with the reader's verdict and does not weaken it, I recommend no change to the REJECT verdict.","tokens_in":23136,"tokens_out":6342,"duration_ms":69630,"concrete_test":"Inspect the promised open-source repository (the Availability section cites a GitHub repository but omits the URL): locate the generator-loss update in the training loop and check whether the L1 term for G compares Q(G(x)) with x as in Eq. (4), or with y as printed in Algorithm 1 line 22, and likewise for F. If the code is unavailable, request it from the authors and re-run the label-0 Edge Detection experiment. If the implemented loss matches Algorithm 1 rather than Eqs. (4)–(5), or if the results change materially when the loss is corrected to Eqs. (4)–(5), the central empirical claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that iHQGAN performs unpaired I2I translation by enforcing content consistency through the unidirectional cycle constraints Q(G(x)) ≈ x and R(F(y)) ≈ y, as defined in Eqs. (4)–(5). However, the training pseudocode in Algorithm 1 prints, for the G update, η||Q(G(x)) − y||₁ (line 22) and, for the F update, η||R(F(y)) − x||₁ (line 28). These swap the reconstruction target: the first enforces that the ACNN output match the target-domain sample rather than the source x, and the second enforces match to x rather than y. The generator loss therefore is not the one in Eqs. (8)–(9), and the reported FID/SSIM/PSNR numbers cannot be cleanly attributed to the claimed cycle-consistency mechanism. This is a structural inconsistency in the only implementation artifact provided; the manuscript states code is open-sourced but gives no link, so a reader cannot resolve the ambiguity by inspection. The issue is load-bearing because the unidirectional cycle constraint is the mechanism that supposedly preserves content, and all headline comparisons depend on it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes iHQGAN, a hybrid quantum-classical generative adversarial network for unpaired image-to-image (I2I) translation. Two quantum generators G and F are constructed as mutually inverse parameter-shared circuits, each paired with a classical \"assisted\" network (ACNN) to enforce a unidirectional cycle-consistency constraint. The model is trained with WGAN-GP-style critics and evaluated on 19 MNIST-derived sub-datasets covering edge detection, font style transfer, and image denoising, with FID/SSIM/PSNR comparisons against low-complexity CycleGAN, One2One, and a Gaussian filter. The central claims are that iHQGAN is the first versatile quantum method for unsupervised I2I translation and that it can outperform classical low-complexity generators while reducing parameter count through quantum reversibility.","tokens_in":23262,"tokens_out":7429,"duration_ms":79325,"significance":"If the claims are supported, the paper would make a useful contribution to quantum machine learning: it applies QGANs to a more complex generation task than previously demonstrated, and the parameter-sharing inverse-circuit design is a concrete and checkable mechanism for reducing generator parameters (5760 versus 35,911×2 for the CycleGAN baseline). The paper also has a broad experimental scope (19 sub-datasets, three tasks) and includes ablations on the cycle-consistency strategy and on loss hyperparameters. These are genuine strengths. However, the empirical component is the main evidence, and the manuscript as submitted has structural inconsistencies in the definition of the training objective, an unspecified component (the ACNN), and comparison-protocol ambiguities. These issues make the reported quantitative results difficult to attribute to the claimed mechanism, so the paper needs a substantial revision before the central claims can be accepted.","major_comments":[{"comment":"The training algorithm is inconsistent with the stated cycle-consistency objective. Equations (4) and (5) define the content-consistency losses as E_x[||Q(G(x)) - x||_1] and E_y[||R(F(y)) - y||_1]. Algorithm 1 line 22 instead prints η||Q(G(x)) - y||_1, and line 28 prints η||R(F(y)) - x||_1, swapping the reconstruction targets. If the printed algorithm is what was run, the model did not enforce the claimed unidirectional cycle constraint; it enforced a cross-domain matching term that is not the content-preservation mechanism described in Section 3.4. Because the manuscript states that code is open-sourced but provides no repository link, the ambiguity cannot be resolved by inspection. This is load-bearing for all comparisons in Tables 5-8, which are attributed to the unidirectional cycle-consistency mechanism.","section":"Algorithm 1, lines 22 and 28; Eqs. (4)-(5)"},{"comment":"The WGAN-GP adversarial loss has a sign inconsistency. Equation (2) is written as E_y[D(y)] - E_y[D(G(x))] - λ E[(||∇_ŷ D(ŷ)||_2 - 1)^2], i.e., real-minus-fake and a minus before the gradient penalty. Algorithm 1 line 10 uses the standard WGAN-GP critic form D(y') - D(y) + λ(||∇_ŷ D(ŷ)||_2 - 1)^2, and line 22 uses the same sign convention for the generator. These two forms are not equivalent and the gradient-penalty term has opposite sign. If Eq. (2) is the intended objective, the training algorithm is using the opposite sign; if the algorithm is correct, Eqs. (2)-(3) are wrong. Since the adversarial term enters the total generator losses in Eqs. (8)-(9), the exact objective actually optimized is left unspecified.","section":"Section 3.5, Eq. (2); Algorithm 1, lines 10 and 22"},{"comment":"The assisted classical neural networks Q and R appear in the objective through Eqs. (4)-(7), but their training is never specified. Algorithm 1 contains no update step for Q or R, and Section 4.2 gives learning rates only for critics, ACNNs, and quantum generators without explaining how the ACNN parameters are optimized. If Q and R are frozen, the manuscript should state what they are and how they were obtained; if they are trainable, the algorithm and implementation details must include their update rule. Without this information, the cycle-consistency loss cannot be evaluated and the reported behavior of iHQGAN cannot be reproduced.","section":"Sections 3.4, 3.5, and Algorithm 1"},{"comment":"The hyperparameters ε=10, η=20, ρ=300 appear to have been selected using test-set performance. Figure 12 is described as a quantitative comparison on the test set, and Section 4.4.3 reports choosing the combination that gives the best FID/SSIM on the sub-dataset with label 0 from the Edge Detection dataset. Selecting model hyperparameters on the same test sets that are subsequently reported in Tables 5-8 can inflate the results and weakens the generalization claims. The authors should either use a validation split for this selection or clearly state that the reported numbers are on held-out test data after a fixed hyperparameter choice made on a separate validation set.","section":"Section 4.4.3 and Figure 12"},{"comment":"The post-processing step described in Section 3.5.2 zeroes out rows 0-7 and 26-31 of the 32×32 iHQGAN outputs, which are border regions that likely contain the discrete noise mentioned in the text. The manuscript does not state whether the same post-processing is applied to CycleGAN, One2One, and Gaussian-filter outputs before computing FID/SSIM/PSNR. If the baselines are not processed identically, the quantitative comparisons in Tables 5-7 are not on equal footing. Please specify the exact evaluation pipeline for every method or apply the same border masking to all outputs.","section":"Section 3.5.2 and Tables 5-7"}],"minor_comments":[{"comment":"The text says that code has been open-sourced, but no repository link is given; Appendix C also refers to a GitHub repository without a URL. Please include the link in the manuscript.","section":"Availability of supporting data"},{"comment":"Equation (5) is labeled L_G_cyc(G,F) but should be L_F_cyc(F,R), and Eq. (3) writes L_DX_GAN(G, DX Y, x, y) with an undefined symbol 'DX Y'. These typos make the loss definitions harder to follow.","section":"Section 3.5 and Algorithm 1"},{"comment":"Two figure references appear as 'Fig. ??'; the schemes Q-wcycleGAN and iHQGAN w/o ACNNs should be given a proper figure number and referred to explicitly.","section":"Section 4.4.2"},{"comment":"The caption for sub-figure (2) contains variable mismatches: it says 'x → G(y) → F(G(y)) ≈ y' where the domain labels are inconsistent with x and y being from different domains. Please correct the caption.","section":"Figure 3"},{"comment":"The implementation details state that the Adam optimizer has 'two weight decay hyperparameters ... set to 0 and 0.9'; this appears to refer to the β1 and β2 momentum parameters. Please use the standard terminology to avoid confusion.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the parameter-sharing inverse-circuit construction is a real strength. However, the manuscript currently has multiple ambiguities in the exact training objective, and the only implementation artifact (Algorithm 1) contradicts the equations that define the central mechanism. Since no code link is provided, a reader cannot determine whether the reported experiments were run with the claimed cycle-consistency loss. These issues are substantive but, in principle, addressable with a careful revision that corrects the equations/algorithm, specifies the ACNN training, equalizes the evaluation protocol, and releases the code. I therefore see this as a major-revision case rather than a clear rejection, provided the authors can resolve the objective ambiguity in a verifiable way."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe core idea here is worth knowing about: iHQGAN is the first attempt I've seen at unsupervised image-to-image translation with a quantum GAN. The trick is to build two generators as mutual inverses by negating and reversing gate parameters of a shared circuit, so the parameter count drops to one circuit. That's a clean use of unitarity, and the 5,760 generator parameters vs CycleGAN's roughly 72k/12k does make the parameter-saving claim concrete. The unidirectional cycle-consistency scheme, where each quantum generator is paired with a classical ACNN that reconstructs the source, is also a sensible way to dodge the weak expressivity of the quantum generator. The experiments span 19 sub-datasets across three tasks, which is more than most QGAN papers bother with.\n\nBut the manuscript in its current form has several load-bearing inconsistencies. Algorithm 1 lines 22 and 28 write the cycle-consistency terms as η||Q(G(x))−y||₁ and η||R(F(y))−x||₁, whereas Eqs. (4)-(5) define them as Q(G(x))≈x and R(F(y))≈y. The SSIM term on the same line uses the correct argument (x for G, y for F), so the pseudocode contradicts itself. Eq. (2) also has a minus sign before the gradient penalty, opposite to both the algorithm and the standard WGAN-GP formulation. Without the promised code (the paper says open-sourced but gives no link), the reader cannot tell which objective was actually optimized. That matters because the unidirectional cycle constraint is the mechanism claimed to preserve content.\n\nThere are additional problems. The headline results on the label-0 Edge Detection sub-dataset come from the same data used to select ε, η, ρ in Section 4.4.3. Post-processing zeroes out rows 0–7 and 26–31 of the generated 32×32 images, and there's no confirmation the classical baselines got the same treatment. The parameter counts in Table 4 omit the ACNN and critic parameters. These are fixable, but each independently weakens the reported FID/SSIM/PSNR comparisons.\n\nAll that said, I don't think this is a crank paper. The conceptual framework is coherent, the related work is cited fairly, and the internal inconsistencies look like sloppy writing rather than deliberate fitting—the theory and architecture hang together. I'd send it to review, but I'd expect the referees to demand a corrected algorithm, consistent equations, a validation split for hyperparameters, and the actual code. If those are provided, the comparative claim can be tested. As is, I wouldn't trust the headline numbers.","headline":"A genuinely new idea—using inverse parameter-shared quantum circuits for unpaired I2I translation—is undermined by internal inconsistencies in the training objective and missing artifacts.","tokens_in":23937,"tokens_out":2463,"would_cite":false,"duration_ms":23763,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Two mutually inverse quantum generators sharing one parameter set, each paired with a small classical helper network enforcing unidirectional cycle consistency, can perform unpaired image-to-image translation and beat low-complexity…","keywords":["quantum machine learning","quantum generative adversarial network","unsupervised image-to-image translation","cycle consistency","invertible quantum circuits","parameter-efficient GAN","hybrid quantum-classical model","image denoising"],"falsifier":"Run the published training code and inspect the generator loss lines corresponding to Algorithm 1 lines 22 and 28; if the L1 term is $\\eta\\lVert Q(G(x)) - y\\rVert_1$ rather than $\\eta\\lVert Q(G(x)) - x\\rVert_1$, retrain with the reconstruction targets stated in Equations (4)-(5) and recompute FID, SSIM, and PSNR. If the reported quality does not reproduce, the content-consistency mechanism is not what carried the result.","tokens_in":22801,"feed_emoji":"⚛️","tokens_out":12222,"duration_ms":109060,"temperature":0.7,"pith_summary":"The paper tries to establish that unpaired image-to-image translation—turning, say, an edge sketch into a digit without paired examples—can be done by a hybrid quantum-classical GAN whose two generators are quantum circuits that are each other's inverse. Because quantum circuits are unitary, the backward generator is obtained from the forward one by reversing the gate order and negating the rotation angles, so both directions can share one parameter set. Each quantum generator is paired with a small classical helper network and a unidirectional cycle-consistency loss, which the paper argues is necessary because the quantum generators themselves are not expressive enough for bidirectional cycle consistency. Experiments on 19 MNIST-derived sub-datasets across edge detection, font style transfer, and denoising report better or competitive FID, SSIM, and PSNR than CycleGAN and One2One with low-complexity CNN generators, with a generator parameter count of 5760 versus roughly 35,911 × 2 for CycleGAN. If correct, this extends quantum GANs beyond single-domain image generation and offers a reversible-quantum route to parameter reduction.","feed_headline":"Quantum GAN with inverse circuits beats CNNs on 19 tasks","feed_subtitle":"It shares one parameter set between two reversible quantum generators, cutting GAN parameters while preserving image structure.","key_machinery":"The load-bearing object is the pair of mutually approximately reversible quantum generators with shared parameters, built on the PQWGAN patch architecture. Each of the 32 sub-circuits uses 5 qubits and 12 parameterized blocks of single-qubit rotation gates plus CNOT entanglement; the inverse circuit is formed by negating every rotation angle and reversing gate order, so $u_l = u_k^\\dagger$. This construction is what lets one parameter set serve both translation directions. The second mechanism is the unidirectional cycle-consistency constraint implemented through the assisted classical networks: because the quantum generators alone have limited expressiveness, bidirectional consistency between two quantum generators fails, but a classical helper network that maps $G(x)$ back toward $x$ and $F(y)$ back toward $y$ steers the quantum parameter space and adds structure-preserving gradients.","core_discovery":"The central claim is that unsupervised image-to-image translation is an approximately reversible task, and quantum circuits are naturally invertible, so a GAN for this task can be built from two quantum generators that are mutual inverses and share their parameters. Writing the forward generator's $k$-th sub-circuit as $u_k$, the backward generator uses $u_l = u_k^\\dagger$, obtained by negating the gate parameters and reversing the gate order; when $k=l$, $u_k u_l = I$, so content encoded in the amplitude state can in principle be recovered. To keep the generated image close to the source content, the model adds an assisted classical neural network after each quantum generator and enforces unidirectional cycle losses $E\\lVert Q(G(x)) - x\\rVert_1$ and $E\\lVert R(F(y)) - y\\rVert_1$, together with an SSIM-based quality-aware loss and Wasserstein adversarial losses. On 19 MNIST-derived sub-datasets spanning edge detection, font style transfer, and denoising, the paper reports that iHQGAN produces smoother, structurally consistent images and that most best FID, SSIM, and PSNR values fall on iHQGAN, while its two shared quantum generators use one parameter set instead of CycleGAN's two.","pith_inferences":["The parameter comparison in Table 4 counts generator parameters only (5760 for iHQGAN versus 6157×2 and 35911×2 for the CNN baselines); if the two assisted classical networks' parameters were included, the advertised saving would be smaller, and a fair same-scale comparison should report them.","A natural next experiment is to push iHQGAN beyond 32×32 single-channel images; amplitude encoding into 5 qubits ties the current representational power to tiny patches, so the versatility claim depends on a circuit design that scales.","The training pseudocode discrepancy matters: Equations (4)-(5) define reconstruction toward the source domain, but Algorithm 1 lines 22 and 28 print cross-domain L1 terms; if the code follows the pseudocode, the reported quality may come from a different objective than the one the paper motivates, and rerunning with the intended objective would settle whether the mechanism is responsible.","A testable extension is to replace the quantum inverse circuits with classical reversible generators of the same parameter budget, which would isolate whether quantum unitarity, rather than reversibility per se, explains the improvement."],"forward_implications":["iHQGAN would be the first quantum GAN formulation for unsupervised image-to-image translation, moving QGAN research from generating single images to learning mappings between two visual domains.","The parameter saving follows from reversibility alone: two mutually inverse circuits share one parameter set, so the generator side costs the same as a single generator rather than two.","Bidirectional cycle consistency between quantum generators is not a free upgrade: comparisons with the two quantum schemes indicate that limited-expressiveness quantum generators need the classical helper networks to keep both directions consistent.","On small grayscale tasks, a quantum generator can beat low-complexity CNN generators on FID, SSIM, and PSNR, so quantum generators need not be confined to toy distributions.","The same reversible mechanism transfers to other approximately reversible translation tasks, and the paper's denoising results are offered as evidence of generalization beyond the two main tasks."],"supporting_citations":[{"why":"Supplies the PQWGAN patch-based quantum generator architecture and critic design on which iHQGAN is built.","marker":"Tsang et al. (2023)"},{"why":"Defines unpaired image-to-image translation with cycle-consistent adversarial networks; CycleGAN is the main classical irreversible baseline.","marker":"Zhu et al. (2017)"},{"why":"Provides the One2One reversible baseline and the self-inverse parameter-sharing idea iHQGAN compares against.","marker":"Shen et al. (2020)"},{"why":"Introduces the quality-aware loss that iHQGAN adapts with SSIM to preserve structural consistency.","marker":"Chen et al. (2019)"},{"why":"Introduces quantum generative adversarial networks and the principle that the generator should be quantum.","marker":"Lloyd and Weedbrook (2018)"},{"why":"Shows a patch strategy for quantum image generation with limited qubits, which PQWGAN and iHQGAN extend.","marker":"Huang et al. (2021)"}],"fun_headline_variants":["Quantum GAN's reversible circuits slash parameters, beat CNNs","Invertible quantum GAN does image translation with shared weights","iHQGAN: quantum invertibility cuts GAN cost, wins on 19 tasks","Quantum GAN uses mutual inverse generators for translation","Sharing inverse circuits: quantum GAN outperforms CNN rivals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results assume the implementation minimized the stated reconstruction losses, in which translating an image forward and then back returns it to its original domain; the pseudocode's generator updates instead print losses that compare the reconstruction to the opposite domain, so the model may have been trained on a different objective than the one the paper claims.","fun_headline_variants_meta":{"raw":{"variants":["Quantum GAN's reversible circuits slash parameters, beat CNNs","Invertible quantum GAN does image translation with shared weights","iHQGAN: quantum invertibility cuts GAN cost, wins on 19 tasks","Quantum GAN uses mutual inverse generators for translation","Sharing inverse circuits: quantum GAN outperforms CNN rivals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1389,"prompt_tokens":1093,"completion_tokens":296,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":709,"completion_tokens_details":{"reasoning_tokens":208}},"tokens_in":709,"tokens_out":296,"duration_ms":3311,"temperature":1.0,"reasoning_tokens":208,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:44:19.443569+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the published training code and inspect the generator loss lines corresponding to Algorithm 1 lines 22 and 28; if the L1 term is $\\eta\\lVert Q(G(x)) - y\\rVert_1$ rather than $\\eta\\lVert Q(G(x)) - x\\rVert_1$, retrain with the reconstruction targets stated in Equations (4)-(5) and recompute FID, SSIM, and PSNR. If the reported quality does not reproduce, the content-consistency mechanism is not what carried the result.","supporting_citations":[],"review_version":1}