{"id":"9a8d18da-1dc1-48a6-bdc3-cdfee1338331","arxiv_id":"2603.00233","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A single end-to-end quantum generator using an image-tailored circuit and learnable multimodal noise achieves state-of-the-art simulated FID scores on full MNIST and Fashion-MNIST without tricks.","lead":"Quantum machine-learning researchers trained a single quantum circuit to generate full 32×32 images covering all ten MNIST and Fashion-MNIST classes, avoiding the usual tricks of shrinking images or stitching patches together. It matters because it offers a concrete recipe for scaling quantum generative models, though only in simulation and with no claimed advantage over classical models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fashion-MNIST comparison uses manually selected modes from a 10-class model rather than a model trained on the 2-class subset, so the claimed 60 vs 179 FID improvement is not a valid head-to-head benchmark.","rationale":"The reader's weakest_assumption centered on FID comparability across resolutions and FID limitations on low-resolution grayscale data. That is a valid concern, but the more directly falsifiable flaw is the non-equivalent benchmarking protocol for Fashion-MNIST: the proposed model was not trained on the 2-class subset, and the evaluated samples were selected post hoc from a 10-class model by visual inspection. This makes the 60-vs-179 comparison unsuited to support the 'new state-of-the-art' claim as stated. The MNIST 3-class comparison is fairer and still provides some support for an improved quantum baseline. Neither concern invalidates the core qualitative contribution—a single end-to-end QGAN generating diverse full-resolution images in simulation—nor the value of the code release. Therefore the appropriate disposition remains conditional acceptance, with the requirement that the Fashion-MNIST benchmark be rerun under a matched training protocol (and that FID preprocessing be specified). The reader already reached CONDITIONAL; my concern reinforces that judgement without moving it.","tokens_in":27296,"tokens_out":7662,"duration_ms":76002,"concrete_test":"Retrain the proposed QGAN from scratch on the 2-class Fashion-MNIST subset (t-shirt, trousers) using the same configuration and training protocol as the 3-class MNIST benchmark (e.g., 16 layers, 2 modes, 15,000 generator updates; exact settings in App. A.4). Generate 10,000 samples and compute FID against the real 2-class subset following the same preprocessing and reference data used for the patch-QGAN baseline. If this FID is not below the reproduced patch-QGAN's 179, then the claim of substantially outperforming the baseline on Fashion-MNIST is not supported, and the SOTA claim must be restricted to MNIST.","verdict_should_be":"UNCHANGED","load_bearing_attack":"App. B7 states: 'we did not retrain a QGAN model in our framework on the 2-class subset of Fashion-MNIST. Instead, we present only samples ... from the modes corresponding to the two classes t-shirt and trousers, which is clear through visual inspection.' This is a post-hoc selection from a model trained on a different task (10 classes), not a model trained on the benchmark subset. The reproduced patch-QGAN baseline (Tsang et al.) was trained specifically on the 2-class Fashion-MNIST subset. FID is computed on 10,000 samples from those manually chosen modes. This breaks the symmetry of the comparison: the baseline was not allowed to select its best outputs, while the proposed method is evaluated on a cherry-picked subset of its output distribution. Even if the authors argue the selection may not advantage them, the protocol does not support the quantitative claim 'our model reaches an FID of 60, substantially outperforming the patch-QGAN result of 179' (Sec. V). The MNIST 3-class comparison (152 vs 207) is fair and supports a weaker SOTA claim, but the Fashion-MNIST number cannot be used as presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an end-to-end quantum Wasserstein GAN for full-resolution image generation. The generator is a parameterized quantum circuit tailored to the FRQI (and MCRQI) image encoding, using a Morton-ordered address-qubit layout, alternating nearest- and next-nearest-neighbor entangling ladders, and controlled color-qubit rotations. A tuned multimodal Gaussian-mixture noise input is introduced to increase intra-class diversity. Experiments are carried out in numerical simulation on MNIST, Fashion-MNIST, and a class-restricted SVHN. The authors report FID values of 118 on 10-class MNIST, 91 on 10-class Fashion-MNIST, and 152 versus 207 on 3-class MNIST when compared with a reproduced patch-QGAN baseline, as well as 60 versus 179 on a 2-class Fashion-MNIST comparison. They also present ablations supporting the task-specific ansatz and tuned multimodal noise, an analysis of shot-noise training, and empirical gradient-scaling evidence relevant to barren-plateau concerns.","tokens_in":27537,"tokens_out":2911,"duration_ms":29461,"significance":"If the quantitative claims hold, this would be a notable step for quantum generative modeling: a single quantum generator, without PCA/autoencoder latents or patch-based multi-generator setups, producing recognizable full-resolution samples on standard datasets. The paper is strengthened by several good practices: the patch-QGAN baseline is reproduced from the authors' released code rather than quoted; the model-selection protocol via MMD is automated and described; the code and data are publicly available; and the ablation study cleanly separates the effect of the task-specific ansatz from the FRQI encoding. The visual samples and ablations provide credible evidence that the architectural inductive bias and multimodal noise are the main drivers of the reported behavior. The main weakness is that the quantitative state-of-the-art claim rests on FID comparisons that are partly not apples-to-apples, and the reported FIDs are single-run point estimates without uncertainty quantification.","major_comments":[{"comment":"The Fashion-MNIST comparison (FID 60 vs 179) is not a valid head-to-head benchmark. App. B7 states that no QGAN was retrained on the 2-class Fashion-MNIST subset; instead, samples were taken from manually identified modes of a 10-class model corresponding to t-shirt and trousers. The reproduced patch-QGAN baseline was trained specifically on the 2-class subset. This asymmetry means the baseline was not allowed to select its best modes while the proposed method was evaluated on a post-hoc selected subset of its output distribution. The sentence in App. B7 claiming this 'should not give our QGAN any advantage' is not a substitute for a symmetric protocol. This comparison is load-bearing for the main 'substantially outperforming' claim in Sec. V, so it should either be removed from the quantitative claims or replaced by a model retrained on the same 2-class subset.","section":"Sec. V / App. B7"},{"comment":"The FID comparisons may be confounded by a resolution mismatch. The proposed QGAN generates 32x32 images (after bilinear interpolation of the datasets), while the reproduced patch-QGAN baseline is explicitly trained and evaluated at the original 28x28 resolution. The benchmarking paragraph in App. B7 does not state whether FID inputs were resampled to a common resolution before Inception-v3 feature extraction. App. B1 acknowledges known FID sensitivity to resizing/aliasing artifacts (citing Ref. [83]). Since the main quantitative claims are 152 vs 207 and 60 vs 179, the paper must either specify the exact preprocessing applied to both sets of images before FID computation or add an analysis showing that the reported differences are robust to resolution normalization.","section":"App. B7 / Sec. IV"},{"comment":"All FID values in the paper appear to be single-run point estimates. GAN training is stochastic, and the largest models are selected via an MMD-based checkpoint criterion; the reported FID therefore has multiple sources of variability (initialization, data shuffling, checkpoint selection, finite sample size). No error bars, number of seeds, or statistical significance statements are provided. This is particularly important for the comparisons against the reproduced baseline, where differences of 55 (152 vs 207) and 119 (60 vs 179) are presented as conclusive. The authors should report at least a small number of independent runs, or alternatively state clearly that the numbers are single-seed observations and soften the corresponding claims.","section":"Sec. IV / App. B2"},{"comment":"The paper appropriately notes in App. B1 that FID is 'potentially less well suited' for MNIST/Fashion-MNIST because the Inception network is trained on much larger natural color images. However, the abstract and Sec. V still make a 'new state-of-the-art performance' claim based on FID. Given the acknowledged metric limitations, the quantitative claim should be accompanied by complementary evidence that the FID gap is meaningful for these datasets, or the claim should be explicitly framed as provisional. This is not a request to remove FID, but to align the strength of the wording with the caveats already present in the appendix.","section":"Sec. V / App. B1"}],"minor_comments":[{"comment":"Typo: 'task-sepcific' should be 'task-specific'.","section":"App. A.1 heading"},{"comment":"Typo: 'path-based' should be 'patch-based' in the comparison paragraph.","section":"App. B7 / Fig. 16"},{"comment":"The caption states that images are 'manually ordered and, where relevant, matched to classes.' This is acceptable for visualization but should be stated even more prominently in the main text so that readers do not interpret the gallery as random samples. The sample-selection rule 'closest to the mean of 500 samples' should also be described before Fig. 3 appears, not only in Sec. IV.","section":"Sec. IV A / Fig. 3"},{"comment":"The notation z_{m,l,a} is used before the layer index l is introduced in the surrounding text. A sentence defining the index ranges of l and a would improve readability.","section":"Sec. III / Eq. (5)"},{"comment":"The sentence 'This encoding is attractive because it allows for representing an image with 2^A pixels using only A qubits' is clear, but it would be useful to explicitly note that the global normalization removes intensity scaling, since this becomes relevant to the later comparison with FRQI.","section":"Sec. II A"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper is worth your time: it is the first QGAN I know that trains a single end-to-end quantum generator on all ten MNIST/Fashion-MNIST classes at 32×32 in simulation, with code and a reproduced baseline rather than borrowed numbers. The architecture story is the real contribution: an FRQI-tailored ansatz plus learnable multimodal noise. The ablation (task-agnostic vs task-specific, amplitude vs FRQI) shows the inductive bias matters, and the layer-wise entropy analysis gives a plausible mechanism. That part is solid and novel.\n\nThe soft spots are quantitative. The claimed \"new state-of-the-art\" rests on three FID comparisons. The MNIST 3-class comparison (152 vs 207) is fair, since they reproduce Tsang et al. and compare at the same protocol. The Fashion-MNIST comparison (60 vs 179) is not: the baseline was trained on the 2-class subset, while their number comes from manually selecting the t-shirt/trouser modes of a 10-class model. That is post-hoc selection from a different task. The authors say it should not disadvantage them—and they may even be right that mixing with other classes hurts rather than helps—but \"may be right\" is not a benchmark. As written, that number overstates the result. The paper itself flags that they did not retrain, so this is not a hidden flaw; it is a design choice that needs a retrained 2-class model or a downgraded claim.\n\nAlso, all FIDs are single-run point values with no seeds, and the 28×28 vs 32×32 resolution mismatch is not addressed in the FID computation. Those two issues are fixable and do not shake the qualitative claim, but they do shake \"new state-of-the-art.\" The authors know FID is ill-suited for MNIST/Fashion-MNIST (App. B1) and still lean on it heavily.\n\nNet: I would send this to referees. The core empirical finding—end-to-end quantum image generation at full resolution, with the architecture choice as the lever—is meaningful and believable. It needs revision: averaged metrics, explicit FID resolution handling, and a fair Fashion-MNIST benchmark. The paper is honest enough that the flaws are visible without a microscope.\n\nA good reading-group paper for QML folks. I would cite it for the FRQI-tailored ansatz result, not for the SOTA numbers.\n\nRecommendation: engage, conditional accept path.","headline":"Strong empirical result with a genuine architecture insight, but the Fashion-MNIST SOTA comparison is not a valid head-to-head benchmark, and the FID claims need seed-averaged, resolution-matched numbers.","tokens_in":28083,"tokens_out":1920,"would_cite":true,"duration_ms":19789,"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":"A single quantum generator designed around the image encoding produces full-resolution, class-diverse MNIST and Fashion-MNIST images in simulation, without dimensionality reduction or patch tricks.","keywords":["quantum generative modeling","quantum Wasserstein GAN","image generation","FRQI encoding","inductive bias","multimodal noise","shot noise","MNIST"],"falsifier":"A decisive check: recompute FID after resampling all real and generated images—patch-QGAN 28×28 outputs, new 32×32 outputs, and real references—to a common resolution (for instance 32×32 or 28×28) and verify that the reported gaps (152 vs 207 and 60 vs 179) persist; if they shrink or reverse, the state-of-the-art claim is not established.","tokens_in":27138,"feed_emoji":"⚛️","tokens_out":9666,"duration_ms":90681,"temperature":0.7,"pith_summary":"The paper argues that quantum image generation has been held back by generic circuit designs, not by quantum hardware. It builds a quantum Wasserstein GAN whose generator circuit is deliberately matched to the FRQI image encoding—one color qubit plus a ladder of spatial entangling gates—and feeds it with learnable multimodal noise. In numerical simulation this single end-to-end generator synthesizes coherent 32×32 images across all ten MNIST and Fashion-MNIST classes, and a color extension handles SVHN. Reported FID scores (118 for MNIST, 91 for Fashion-MNIST) beat a reproduced patch-based quantum baseline (152 vs 207 on 3-class MNIST; 60 vs 179 on 2-class Fashion-MNIST) with no PCA/autoencoder latent space and no per-patch generators. The takeaway is that task-specific inductive bias, including the noise distribution, is the lever that makes quantum generative models scale to standard benchmark datasets.","feed_headline":"One quantum circuit renders all ten MNIST digit classes","feed_subtitle":"The same single generator also covers Fashion-MNIST and colored SVHN, beating a reproduced patch-based quantum baseline.","key_machinery":"The argument runs on two mechanisms. First, the FRQI (flexible representation of quantum images) encoding stores a 2^A-pixel image in A address qubits plus one color qubit, and the generator circuit is designed to bias outputs toward valid FRQI states: Hadamards prepare a uniform gray state, Rx gates upload noise, alternating nearest-neighbor and next-nearest-neighbor entangling ladders on address qubits mix spatial scales along Morton order, and controlled Ry rotations set pixel intensities. Second, learnable multimodal noise—a uniform Gaussian mixture with trainable means and variances—provides the latent-space diversity that prevents class blending and supports intra-class variation. A de","core_discovery":"The central discovery is that a quantum generator can learn realistic low-resolution images directly if its variational circuit is biased toward the structure of the image encoding. The proposed generator prepares FRQI states—pixel intensities stored in a color qubit whose address-qubit controls follow Morton order—using alternating nearest-neighbor and next-nearest-neighbor entangling gates that mix spatial dimensions at increasing scales, followed by controlled color rotations. This architecture, paired with a Gaussian-mixture latent space whose per-mode means and variances are trainable, produces all ten classes of handwritten digits and clothing items with visible intra-class variety. Th","pith_inferences":["A testable extension of the inductive-bias explanation: if the ansatz–encoding match is the true cause of the gains, the ablation gap between task-specific and task-agnostic circuits should grow with image size; training at 64×64 and 128×128 would show a widening FID gap.","Given the paper's own caution that FID is 'potentially less well suited' for these datasets, we would expect a resolution-matched or distribution-matched evaluation (e.g., resampling all images to a common size) to alter the absolute FID values, even if the qualitative ordering of quantum models survives.","The shot-noise smoothing result suggests a broader regularizer principle for amplitude-type encodings; a direct transfer experiment would be to train the same QGAN on time-series or tabular amplitude-encoded data with and without finite-shot noise to see whether the smoothing effect suppresses memorization.","The layer-wise entropy analysis indicates a coarse-to-fine generation pathway; if so, a curriculum that trains early layers first and fine-tunes only the fine-resolution ladders later should cut training cost, which could be tested by comparing convergence curves."],"forward_implications":["Quantum image generation on standard benchmark datasets no longer requires dimensionality reduction or multiple patch generators; a single quantum generator suffices for at least 32×32 grayscale and color images, moving the field beyond toy examples.","The ansatz's spatial entanglement ladder mirrors the Morton-order pixel hierarchy, so the same principle—matching the circuit to the data's hierarchical compression—should carry over to other structured data loadable via similar encodings.","Learnable multimodal noise acts as a direct counter to mode collapse: with more modes than classes, the model separates sub-styles within a class and even produces sensible inter-class blends, suggesting a general recipe for diversity in quantum generative models.","Training with finite shot noise can improve generated image quality relative to exact simulation, and the required shot budget scales with pixel count rather than exponentially with qubit count, which is encouraging for physical implementation on early fault-tolerant devices."],"fun_headline_variants":["Single quantum circuit for all ten digit classes","Quantum GAN does full MNIST without image tricks","One quantum generator, ten classes, zero tricks","Full-resolution quantum image gen without shortcuts","Quantum model scales to complete MNIST and more"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the FID numbers are trustworthy and comparable for the comparison made—specifically, that FID computed on 32×32 generated images and 28×28 baseline images of the same low-resolution grayscale datasets is a fair yardstick even though the paper itself notes FID is 'potentially less well suited' for them.","fun_headline_variants_meta":{"raw":{"variants":["Single quantum circuit for all ten digit classes","Quantum GAN does full MNIST without image tricks","One quantum generator, ten classes, zero tricks","Full-resolution quantum image gen without shortcuts","Quantum model scales to complete MNIST and more"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00032,"raw_usage":{"total_tokens":1632,"prompt_tokens":728,"completion_tokens":904,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":472,"completion_tokens_details":{"reasoning_tokens":835}},"tokens_in":472,"tokens_out":904,"duration_ms":8773,"temperature":1.0,"reasoning_tokens":835,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T19:59:36.527574+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check: recompute FID after resampling all real and generated images—patch-QGAN 28×28 outputs, new 32×32 outputs, and real references—to a common resolution (for instance 32×32 or 28×28) and verify that the reported gaps (152 vs 207 and 60 vs 179) persist; if they shrink or reverse, the state-of-the-art claim is not established.","supporting_citations":[],"review_version":1}