{"id":"6d2565c1-34fb-44d8-ace8-5660d5dddb3e","arxiv_id":"2504.17804","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A spectral dictionary model that reconstructs images as linear combinations of learned sinusoidal bases with a Gaussian prior on coefficients is proposed, but the paper does not provide a workable algorithm or reproducible evidence.","lead":"This paper proposes to generate CIFAR-10 images by learning a global dictionary of sinusoidal basis functions and sampling per-image mixing coefficients from a fitted Gaussian. The authors claim competitive FID/IS scores, but the manuscript omits core implementation details, algorithm specifications, and reproducible experimental protocols.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The generative claim is unsupported by the text: Section 2 never specifies how per-image mixing coefficients are obtained, and since Eq. (1) is linear in w, a Gaussian prior makes generated images Gaussian in pixel space, which is not shown to be competitive on CIFAR-10.","rationale":"The paper's own text supports the objection at multiple points. Section 2.2 never writes the per-image coefficient optimization; Figure 1's caption is the only mention of an encoder, and it is vague ('encoder or sparse coding module'). Section 2.3 describes fitting a multivariate Gaussian to coefficients, but if coefficients cannot be computed, the generative pipeline cannot be instantiated. Additionally, because Eq. (1) is linear in w, a Gaussian prior implies the generated distribution is Gaussian in pixel space; the paper provides no evidence that this distribution is close to CIFAR-10. The reported FID 55.4 and IS 7.2 are not reproduced, and there are no sample images. In a generous reading, the method is an under-specified linear autoencoder with a Gaussian latent, close to probabilistic PCA, so the novelty and performance claims are not supported. The reader identified the same fragile assumptions, though I weight the unspecified coefficient estimator at least as heavily as the unimodality of the prior, hence partial agreement. My concrete check, reimplementing with a least-squares or specified encoder and measuring FID/IS, would settle whether the table's numbers and the 'competitive' claim are real.","tokens_in":8121,"tokens_out":6391,"duration_ms":63017,"concrete_test":"Obtain the missing coefficient-estimation specification; if unavailable, reimplement the closest complete pipeline from the text, e.g., estimate w as the pseudo-inverse of S applied to x at fixed dictionary or with the learned modulation network, fit a multivariate Gaussian to training coefficients, draw 50,000 samples, and compute FID/IS against the CIFAR-10 training set with a standard Inception backbone. Compare with Table 1; if FID/IS cannot be reproduced, or if generated samples collapse toward a mean/linear subspace with visible blur, the central generative claim fails.","verdict_should_be":"REJECT","load_bearing_attack":"Section 2.3 states that after jointly optimizing the dictionary and per-image mixing coefficients {w^(n)}, a simple prior such as a multivariate Gaussian is fitted to these coefficients, and new images are generated by sampling w* and forming xhat*(t) = sum_i w_i^* s_i(t). Two linked gaps block this central claim. First, the coefficients w^(n) are never actually defined. Section 2.2 gives only loss functions; there is no optimization problem for w, no update rule, and no architecture for the 'encoder or sparse coding module' that appears only in the Figure 1 caption, which itself offers two incompatible options. Without a rule that maps x to w, the prior cannot be fitted and the entire Table 1 is unreproducible. Second, even granting an estimator, the generation step is a linear map w -> S w. With w ~ N(mu, Sigma), xhat* is Gaussian in pixel space with mean S mu and covariance S Sigma S^T. A unimodal Gaussian over a K-dimensional coefficient space cannot represent the multimodal, class-structured CIFAR-10 distribution; it will place high probability near the mean dictionary combination. The paper supplies no generated images, no per-class analysis, no error bars, and no code to substantiate the claim that this linear Gaussian pipeline achieves FID 55.4 and IS 7.2, or that those numbers are competitive. The reader's REJECT verdict is supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a generative model for images based on spectral dictionary learning. Each CIFAR-10 image is flattened to a 3072-dimensional signal and approximated as a linear combination of K learned basis functions, each a sinusoid with time-modulated amplitude, frequency, and phase (Eqs. 1-5). Training minimizes a composite loss combining time-domain MSE and an STFT-based spectral magnitude loss (Eqs. 6-8). After training, a simple probabilistic prior, e.g., a multivariate Gaussian, is fit to the per-image mixing coefficients, and new images are generated by sampling coefficients and applying the fixed linear dictionary. The paper reports FID 55.4 and Inception Score 7.2 on CIFAR-10 and claims competitive performance, interpretability, training stability, and efficiency relative to VAEs, GANs, and diffusion models.","tokens_in":8474,"tokens_out":8237,"duration_ms":71155,"significance":"The high-level idea of replacing stochastic deep generative models with an explicitly parameterized, interpretable spectral dictionary plus a fitted prior is refreshing and, if fully implemented and validated, could be a useful contribution to interpretable and controllable synthesis. The equation-level description of the basis functions and the two-stage reconstruction-plus-prior scheme is clear. However, the manuscript as it stands does not establish the central claims: the coefficient estimation step is unspecified, the Gaussian-prior generation mechanism is theoretically expected to collapse to a unimodal blur, and the reported numbers are not reproducible from the given text. No code, generated images, hyperparameters, or error bars are provided. The contribution is therefore currently an outline rather than a substantiated method.","major_comments":[{"comment":"The per-image mixing coefficients w^(n) are never defined or computed. Section 2.2 gives only the losses in Eqs. (6)-(8), with no optimization problem, update rule, or architecture for producing w from x; the only reference to an 'encoder or sparse coding module' appears in the Figure 1 caption, and the two options are left unspecified. This gap is load-bearing: without a rule mapping each image to a coefficient vector, the prior p(w) in Section 2.3 cannot be fitted, the generation pipeline cannot be executed, and Table 1 cannot be reproduced.","section":"Section 2.2 / Section 2.3 / Figure 1"},{"comment":"The generative mechanism is a fixed linear map: after training, sampled coefficients w* produce \\hat x* = S w*, where S is the dictionary. If w* is drawn from a multivariate Gaussian, as the paper explicitly suggests, then \\hat x* is Gaussian in pixel space with mean S mu and covariance S Sigma S^T. A single unimodal Gaussian over coefficient space cannot represent the multimodal, class-structured CIFAR-10 distribution, and the expected samples are near the mean dictionary combination, i.e., blurry or collapsed. The paper provides no generated images, per-class diversity statistics, or sample-quality analysis to show that this is not what happens, so the claim of competitive generative performance is not supported.","section":"Section 2.3"},{"comment":"The experimental section omits all details required to interpret or reproduce the results: the number of bases K, the STFT loss weight lambda_STFT, the modulation-network architecture, the STFT window/hop/nfft parameters, the number of samples used for FID and Inception Score, and any error bars. It is also unclear whether the FID/IS evaluation used a held-out split or the training set, since the text only says 'CIFAR-10 benchmark.' The baseline numbers are not documented: the diffusion-model citation [20] is the original 2015 paper, which did not report FID 50.0 or IS 7.5 on CIFAR-10, and the VAE and GAN entries are not tied to specific implementations. The table is therefore not a controlled comparison, and the abstract's claim of 'competitive performance' is not established.","section":"Section 3 / Table 1"},{"comment":"The qualitative claims are unverifiable because no generated or reconstructed image is shown. The text states that 'the generated images exhibit visual fidelity,' but Figure 2 is only a heatmap of mixing coefficients, and the paper's own 'Discussion on Heatmap Realism' concedes that this heatmap consists of uniform color bands and 'might not convey the nuanced information' of a useful attention map. Without sample images, per-class examples, or a visualization of the learned dictionary atoms, the interpretability and visual-quality claims rest entirely on assertion.","section":"Section 3 / Figures 1-2"}],"minor_comments":[{"comment":"The signal is defined on the continuous interval t in [0,1] after flattening a discrete 32x32x3 image, but the discretization, ordering of color channels, and resampling steps are not specified; Eqs. (1)-(5) should be stated for discrete time.","section":"Section 2.1"},{"comment":"The phrase 'carefully tuned hyperparameters' is never followed by actual values; please report K, lambda_STFT, learning rate, batch size, and the modulation-network architecture.","section":"Section 3"},{"comment":"The manuscript references Figure 2 but does not include the figure; the caption and the 'Discussion on Heatmap Realism' should be reconciled or removed.","section":"Figure 2"},{"comment":"Reference [20] is mis-cited: the author list should be Sohl-Dickstein, Weiss, Maheswaranathan, and Ganguli, and that paper does not contain the CIFAR-10 numbers shown in Table 1.","section":"References"},{"comment":"The abstract and Section 4 call the generation process 'deterministic,' but it begins by sampling w* ~ p(w); the writing should say synthesis is deterministic after sampling.","section":"Abstract / Section 4"},{"comment":"There are numerous typographical issues, including the unrendered 'KX' before the sum in Eq. (1), the spacing in 'V AEs,' and stray spaces before colons in Section 4 bullet lists.","section":"Throughout"},{"comment":"The paper claims improved training stability and computational efficiency but provides no runtime, convergence curves, or parameter-count comparison against the baselines.","section":"Discussion"}],"recommendation":"reject","confidential_remarks":"To the editor: the manuscript appears to be an incomplete draft rather than a finished paper. The central algorithmic step is missing, the diffusion baseline in Table 1 is cited to the wrong paper, the only qualitative figure is acknowledged in the text to be uninformative, and no code or samples are supplied. I would not recommend inviting a resubmission in its current form; a viable version would need a fully specified coefficient estimator, real generated images, error bars, and a proper comparison. The spectral dictionary idea itself may be worth exploring, but the present paper does not demonstrate it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Andrew, quick take on arXiv:2504.17804. The short version: this is a rough draft, not a paper. The core model is a linear generative model — Eq. (1) with a Gaussian prior on w is effectively factor analysis / probabilistic PCA with sinusoidal basis atoms. That's not new. The reported results (FID 55.4, IS 7.2) are a single table with no error bars, no generated images, and no implementation details. The paper never says how the per-image mixing coefficients w are computed. Section 2.2 gives losses but no optimization problem for w; the only mention of an encoder or sparse coding module is in the Figure 1 caption. Without that rule, the prior-fitting step in Section 2.3 cannot be executed, and Table 1 is irreproducible.\n\nWhat is worthwhile here: the parametric dictionary idea — sinusoids with time-varying amplitude, frequency, and phase enforced by softplus, plus an STFT magnitude loss — is a reasonable way to build an interpretable, deterministic generator. The math is cleanly written, and the paper is honest about its own limitations in the heatmap discussion. That candor is a point in its favor.\n\nThe soft spots are large. Even granting an encoder, generation is a linear map from w to x, so a Gaussian prior makes x Gaussian in pixel space. That cannot represent the multimodality of CIFAR-10; the gaussian would sit near the mean and produce blurry outputs. No evidence is shown to counter that. The paper reports no hyperparameters (K, λ_STFT), no network architecture for the modulation function, no training curves, and no code. The experimental section is a placeholder. There are also signs of an unedited draft: the self-referential heatmap-realism note, the inflated novelty discussion, and sloppy references.\n\nFor whom is this useful? Possibly as a starting idea in an exploratory sketch, but not as a published claim. I would desk-reject it. It doesn't deserve referee time because the missing method definition blocks any meaningful evaluation. If the authors return with a real implementation, code, samples, and a comparison against PPCA/FA baselines, it might be a modest workshop-level contribution. As is, it's not a paper.","headline":"Under-specified linear generative model with an unreproducible results table; the core idea reduces to probabilistic PCA with sinusoidal atoms, and the reported numbers are unsupported by the text.","tokens_in":8940,"tokens_out":3107,"would_cite":false,"duration_ms":27416,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes a generative image model in which each image is a weighted sum of learned spectral sinusoids, and new images are made by sampling those weights from a fitted Gaussian and applying one linear combination.","keywords":["spectral dictionary learning","generative image modeling","CIFAR-10","short-time Fourier transform","mixing coefficients","interpretable generative model","deterministic synthesis","dictionary learning"],"falsifier":"Fit the model on CIFAR-10, record the per-image mixing vectors, and compare the fitted multivariate Gaussian with the empirical distribution—for example, by checking class-conditional means or Mahalanobis distances. If the coefficients separate into distinct per-class clusters, sampling from the single Gaussian will interpolate between modes and the generated images will be visibly blurry or contain mixed-class content; observing that would falsify the generative claim.","tokens_in":7905,"feed_emoji":"🖼️","tokens_out":12389,"duration_ms":97480,"temperature":0.7,"pith_summary":"This paper tries to show that a classical signal-processing idea—dictionary learning—can serve as a full generative model for images, replacing stochastic latent variables and adversarial training. The model flattens each image into a one-dimensional signal and reconstructs it as a weighted sum of learned spectral atoms whose frequency, phase, and amplitude vary over time. New images are produced by fitting a simple probabilistic prior, such as a multivariate Gaussian, to the per-image weights, then sampling from that prior and taking one linear combination. If this works, image synthesis becomes deterministic, interpretable, and cheap at sampling time, with each atom explicitly tied to a frequency component. The paper reports CIFAR-10 scores of FID 55.4 and Inception Score 7.2, which sit below its VAE baseline and within range of its GAN and diffusion baselines.","feed_headline":"Spectral dictionary generates CIFAR-10 images in one step","feed_subtitle":"Each image is a weighted sum of learned tones; sampling the weights from a simple prior is all it takes to synthesize a new one.","key_machinery":"The carrying object is a global spectral dictionary of $K$ sinusoids with time-varying modulations. Each atom is defined by learned base amplitude, frequency, and phase values plus a modulation network that outputs time-dependent increments, with the smooth positive softplus function keeping amplitude and frequency strictly positive. The representation $\\hat{x}(t)=\\sum_i w_i s_i(t)$ is linear in the per-image weights, so the same dictionary serves both directions: solving for $w$ reconstructs an image, and sampling $w$ from a fitted prior generates one. The short-time Fourier transform magnitude term in the training loss is what forces the reconstructions to preserve the original image's time-frequency content rather than only its pixel values.","core_discovery":"The paper's central claim is that generative image modeling does not need stochastic latent variables or adversarial objectives. It represents each flattened image as $\\hat{x}(t)=\\sum_i w_i s_i(t)$, where each atom $s_i(t)=\\operatorname{softplus}(A_i^0+\\Delta A_i(t)) \\sin(2\\pi \\operatorname{softplus}(f_i^0+\\Delta f_i(t)) t+\\phi_i^0+\\Delta\\phi_i(t))$ is a sinusoid with explicit amplitude, frequency, and phase parameters plus small time-varying modulations learned by a network. The global dictionary and the per-image mixing coefficients $w$ are optimized jointly by minimizing pixel-wise mean squared error together with a short-time Fourier transform magnitude loss. After training, a multivariate Gaussian is fitted to the observed coefficient vectors, and a new image is synthesized by drawing $w^*$ from that prior and computing the same linear superposition. On CIFAR-10 the paper reports FID 55.4 and Inception Score 7.2, below the VAE baseline's 68.0/6.8 and near the diffusion baseline's 50.0/7.5 and GAN baseline's 42.5/8.1.","pith_inferences":["A larger dictionary and a longer training budget than the reported 20 epochs could narrow the FID gap to the GAN baseline, but the paper does not explore this scaling.","The flattened one-dimensional representation discards spatial locality; extending the atoms to be spatially localized or convolutional is a natural next step the paper leaves implicit.","Because the prior is the only stochastic component, replacing the Gaussian with a mixture or a learned density over the weights is a direct test of whether the multimodal structure of natural images is the limiting factor.","The explicit frequency parameters suggest spectral editing—attenuating high-frequency atoms to soften texture, for instance—which the paper motivates but does not demonstrate."],"forward_implications":["New-image synthesis is a single step: draw the weight vector once from the fitted prior and compute one linear combination, so sampling cost is negligible compared with diffusion-style iterative denoising.","Each latent dimension is tied to one spectral atom with explicit frequency, amplitude, and phase parameters, giving a direct handle on frequency content for controlled manipulation of generated images.","Training needs no discriminator, no KL balancing, and no diffusion schedule, so the failure modes the paper attributes to GANs and VAEs—mode collapse and posterior collapse—are structurally avoided.","Because the same dictionary is used for reconstruction and generation, the model doubles as an interpretable autoencoder whose latent space can be inspected through mixing-coefficient visualizations."],"supporting_citations":[{"why":"Establishes the K-SVD dictionary-learning paradigm of jointly refining atoms and codes, which the spectral dictionary adapts.","marker":"[1]"},{"why":"Supplies the classical sparse-coding result that natural image patches yield Gabor-like, frequency-localized atoms, motivating the spectral basis choice.","marker":"[16]"},{"why":"Provides the CIFAR-10 dataset on which reconstruction and generation are evaluated.","marker":"[13]"},{"why":"Supplies the signal-processing background for Fourier representations that the spectral dictionary and STFT loss build on.","marker":"[17]"},{"why":"Serves as the VAE baseline whose reported FID and Inception Score are compared against in Table 1.","marker":"[12]"},{"why":"Serves as the GAN baseline whose reported FID and Inception Score are compared against in Table 1.","marker":"[9]"},{"why":"Serves as the diffusion baseline whose reported FID and Inception Score are compared against in Table 1.","marker":"[20]"},{"why":"Defines the Fréchet Inception Distance used to quantify the quality of generated images.","marker":"[10]"}],"fun_headline_variants":["Images as weighted sinusoids: new generative model","Spectral dictionary: one-step CIFAR-10 generation","Generating images by summing learned tones","No VAEs, GANs, or diffusion: spectral basis wins","Interpretable generative model via frequency dictionary"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole generative step depends on one premise: that a single bell-shaped distribution fitted to the per-image weights preserves enough of the variety of natural images that drawing from it yields new, realistic images rather than blurry averages; the paper also assumes an image-to-weights encoder that it describes only in a figure caption.","fun_headline_variants_meta":{"raw":{"variants":["Images as weighted sinusoids: new generative model","Spectral dictionary: one-step CIFAR-10 generation","Generating images by summing learned tones","No VAEs, GANs, or diffusion: spectral basis wins","Interpretable generative model via frequency dictionary"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000709,"raw_usage":{"total_tokens":3244,"prompt_tokens":1046,"completion_tokens":2198,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":662,"completion_tokens_details":{"reasoning_tokens":2123}},"tokens_in":662,"tokens_out":2198,"duration_ms":13829,"temperature":1.0,"reasoning_tokens":2123,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:40:12.762797+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit the model on CIFAR-10, record the per-image mixing vectors, and compare the fitted multivariate Gaussian with the empirical distribution—for example, by checking class-conditional means or Mahalanobis distances. If the coefficients separate into distinct per-class clusters, sampling from the single Gaussian will interpolate between modes and the generated images will be visibly blurry or contain mixed-class content; observing that would falsify the generative claim.","supporting_citations":[{"cited_title":"Aharon, M","cited_arxiv_id":null,"evidence_quote":"Establishes the K-SVD dictionary-learning paradigm of jointly refining atoms and codes, which the spectral dictionary adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the classical sparse-coding result that natural image patches yield Gabor-like, frequency-localized atoms, motivating the spectral basis choice."},{"cited_title":"Krizhevsky","cited_arxiv_id":null,"evidence_quote":"Provides the CIFAR-10 dataset on which reconstruction and generation are evaluated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the signal-processing background for Fourier representations that the spectral dictionary and STFT loss build on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the VAE baseline whose reported FID and Inception Score are compared against in Table 1."},{"cited_title":"Goodfellow, J","cited_arxiv_id":null,"evidence_quote":"Serves as the GAN baseline whose reported FID and Inception Score are compared against in Table 1."},{"cited_title":"Sohl-Dickstein, E","cited_arxiv_id":null,"evidence_quote":"Serves as the diffusion baseline whose reported FID and Inception Score are compared against in Table 1."},{"cited_title":"Heusel, H","cited_arxiv_id":null,"evidence_quote":"Defines the Fréchet Inception Distance used to quantify the quality of generated images."}],"review_version":1}