{"id":"fea1ab01-25a1-4db7-9514-338408f76df4","arxiv_id":"2505.12552","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"FreqSelect, a learnable band-pass weighting module, yields marginal SSIM and CLIP gains over Brain-Diffuser at N=16 but degrades most other metrics and fails badly at N=8.","lead":"Researchers inserted a trainable frequency filter, FreqSelect, before the image encoder in a two-stage fMRI-to-image reconstruction pipeline. The filter improves two of eight quality metrics in the paper's main table, so the advertised 'consistent improvement' is not supported by the reported numbers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim of consistent improvement is contradicted by Table 1 itself, and the single-subject, no-error-bar results leave no statistically supported win over Brain-Diffuser.","rationale":"The reader and I agree on the weakest assumption: the static global gating contradicts the 'content-aware/dynamic' language, but the decisive problem is simpler and more direct: the paper's own Table 1 fails to show the claimed consistent improvement. For a rejection verdict, the load-bearing concern is that the headline empirical claim is contradicted by the evidence in the same paper. Even granting the static-gate design as a limitation, the paper could have been conditionally acceptable if the numbers had shown a clear win; they do not. The strongest claim (consistent improvement) is falsified on the paper's own numbers, and the absence of error bars and multi-subject quantitative results means we cannot even verify whether the two reported wins (SSIM, CLIP) are real. My independent check is a re-computation of the comparison with subject-level and trial-level statistics; if the gains vanish or are within noise, the reject verdict stands. I therefore agree with the reader and keep the verdict unchanged from the reader's REJECT.","tokens_in":13135,"tokens_out":1477,"duration_ms":12951,"concrete_test":"Ask the authors for the per-subject and per-trial metric tables (or rerun their pipeline) and compute mean±std across the four subjects listed in Section 3.1 for all eight metrics at N=16. If the N=16 gains over Brain-Diffuser on SSIM and CLIP do not replicate on subjects 2, 5, and 7 with overlapping error bars, the 'consistent improvement' claim is not supported. Also recompute Table 1 with a paired bootstrap over the 982 test images; a paired test is needed because the same test stimuli are decoded by both methods.","verdict_should_be":"REJECT","load_bearing_attack":"The abstract and conclusion claim FreqSelect 'consistently improves reconstruction quality across both low- and high-level metrics,' and Section 3.3 claims that at N=16 it 'outperforms Brain-Diffuser on several metrics.' The paper's own Table 1 (subject 1 only, no error bars or significance tests) shows Ours (N=16) beats Brain-Diffuser on exactly two of eight metrics (SSIM 0.2961 vs 0.293; CLIP 92.65% vs 92.5%) and loses on six (PixCorr 0.2734 vs 0.304; AlexNet(2) 96.25% vs 96.84%; AlexNet(5) 97.46% vs 97.48%; Inception 88.36% vs 88.6%; EffNet-B 0.7672 vs 0.761; SwAV 0.4141 vs 0.410). The margin on the two wins is tiny relative to typical run-to-run variability in this pipeline, and the two largest gaps (PixCorr, N=8 EffNet-B) are in FreqSelect's favor only in the wrong direction. Because no standard deviation, per-trial breakdown, or multi-subject table is provided, the central claim is not supported by the reported numbers; the quantitative section itself acknowledges only 'several metrics,' which is inconsistent with 'consistently improves ... across both low- and high-level metrics.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces FreqSelect, a learnable frequency-band weighting module inserted before a frozen VDV AE encoder in a two-stage fMRI-to-image reconstruction pipeline. Each input image is decomposed into N radial frequency bands via DFT masks; scalar weights α_i = σ(w_i) are learned jointly with a ridge regressor that predicts VDV AE latents from fMRI, and the filtered image is a normalized weighted sum of band-limited images. Stage 2 uses the Versatile Diffusion pipeline for refinement. Experiments on the Natural Scenes Dataset (four subjects for qualitative results, subject 1 for the quantitative table) compare N = 4, 8, 16 against Brain-Diffuser and report reconstructions, learned frequency weights, and failure cases.","tokens_in":13446,"tokens_out":8078,"duration_ms":79127,"significance":"If it worked as advertised, FreqSelect would be a simple, cheap plug-in that improves Brain-Diffuser and offers interpretable frequency profiles. The paper is clearly written and the modular design is easy to understand, and the appendix honestly lists limitations and future controls. However, the central quantitative claim is not supported by the authors' own Table 1, and the module is not input-dependent despite being described as content-aware. With a single-subject evaluation and no error bars, the evidence is insufficient to establish the claimed improvements.","major_comments":[{"comment":"The central claim of the abstract and conclusion is that FreqSelect 'consistently improves reconstruction quality across both low- and high-level metrics,' but Table 1 contradicts this. With N = 16, FreqSelect beats Brain-Diffuser on only two of eight metrics (SSIM 0.2961 vs 0.293; CLIP 92.65% vs 92.5%) and loses on PixCorr (0.2734 vs 0.304), AlexNet(2), AlexNet(5), Inception, EfficientNet-B, and SwAV. With N = 4 and N = 8, every metric is worse than Brain-Diffuser. Moreover, results are reported for subject 1 only, without error bars or significance tests, so even the two improvements are not established as reliable. Section 3.3 itself acknowledges only 'several metrics' at N = 16, which is inconsistent with the abstract's 'consistently.'","section":"§3.3, Table 1"},{"comment":"The abstract and Section 1 describe FreqSelect as 'dynamic' and a 'content-aware gate' that 'dynamically adjusts emphasis' based on relevance, but Eq. (5) defines α_i = σ(w_i), scalar parameters that do not depend on the input image x or on the fMRI signal. After training, the same fixed frequency weighting is applied to every image and every subject. If the intended contribution is input-dependent gating, the implementation does not realize it; at most the weights adapt during training, not at inference. This mismatch between the advertised mechanism and the actual formulation needs to be resolved.","section":"§2.1, Eq. (5)"},{"comment":"The abstract claims that 'our method generalizes across subjects and scenes,' but the only quantitative comparison against Brain-Diffuser is for subject 1, as stated in Section 3.3. Figure 6 provides qualitative examples for four subjects, but no subject-wise metrics or statistics are reported. Consequently, cross-subject generalization is not demonstrated by the evidence presented.","section":"§3.3, Table 2; §3.2, Figure 6"},{"comment":"The 'neuroscientific insight' contribution is overstated. The learned frequency weights are optimized to minimize the latent-space MSE between VDV AE latents of filtered images and fMRI predictions; they are therefore a property of the frozen VDV AE encoder, the ridge regressor, and the training objective. Interpreting these weights as 'interpretable evidence about how the brain represents visual information at different spatial scales' requires additional control analyses (e.g., comparing with random features or with a different encoder), which are not provided. This should be reframed as a stimulus-preprocessing analysis rather than direct evidence of cortical frequency tuning.","section":"§3.2, §4, Appendix A.1"}],"minor_comments":[{"comment":"Reference [13] is cited for the claim that Lin et al. first used the Natural Scenes Dataset for fMRI-to-image reconstruction with a StyleGAN2 generator, but the bibliography entry for [13] is a CVPR paper titled 'Deep frequency filtering for domain generalization' and appears unrelated to fMRI decoding. The reported 78.2% Inception value needs a correct source.","section":"References / §3.3"},{"comment":"The 'Inception Score' is described as 'gauging both diversity and classifiability,' but the metric used in this reconstruction benchmark is typically the correlation of Inception feature activations between reconstruction and ground truth; please clarify the exact computation.","section":"§3.1"},{"comment":"The text says the highest-frequency bands 'often reach negative values,' but Eq. (5) defines α_i = σ(w_i) ∈ (0,1); either the figure plots w_i rather than α_i, or the description is inconsistent. Please label the ordinate and clarify.","section":"§3.2, Figure 4"},{"comment":"The sentence before Table 2 says 'Table 2 presents qualitative comparisons,' but it is actually quantitative; also the caption says 'Best scores are in bold' but no bold entries are visible in the text version.","section":"§3.3, Table 2"},{"comment":"There are scattered typos, e.g., 'we applied the General ROI mask' with a lowercase initial, 'Evaluated on the Natural Scenes dataset' with inconsistent capitalization, and 'effNet-B' in the caption.","section":"General"}],"recommendation":"reject","confidential_remarks":"To the editor: the authors may have a useful idea, but the current manuscript's headline claim is contradicted by its own quantitative table, and the described 'content-aware' mechanism is not implemented in Eq. (5). A rejection is appropriate; a future version with per-image gates, multi-subject and multi-seed evaluation, error bars, and honest metric-by-metric claims could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the FreqSelect module is a sensible, easy-to-explain idea, and the paper is clearly written. But the headline claim—'consistently improves reconstruction quality across both low- and high-level metrics'—doesn't survive its own Table 1. At N=16, it beats Brain-Diffuser on SSIM (0.2961 vs 0.293) and CLIP (92.65% vs 92.5%), and loses on the other six metrics. At N=4 and N=8 it's worse across the board. That's not a consistent improvement; it's a marginal gain on two metrics at one setting.\n\nWhat's new here is applying learnable radial frequency-band weighting to fMRI-to-image reconstruction. The method itself is simple: decompose the image into N bands, learn scalar per-band weights, apply a weighted sum before the VDV AE encoder. The integration with the Brain-Diffuser pipeline is clean, and the visualizations of the learned weights (Figure 4) are a nice touch—low frequencies get passed, high frequencies get suppressed, which is exactly what you'd expect from fMRI's low-pass character. The related work section is thorough and honest about prior frequency-gating methods in vision. The paper also openly discusses the N=8 failure and future work, which I appreciate.\n\nSoft spots, in order of severity. First, the central claim is overstated. The abstract and conclusion say 'consistently improves' but the numbers don't support it. The authors do walk it back in Section 3.3 to 'several metrics,' which is accurate but conflicts with the abstract. Second, the quantitative evaluation is on subject 1 only, with no error bars or significance tests. That's not enough to support any claim of improvement, given the known run-to-run variability in this pipeline. Third, the module is described as 'content-aware' and 'dynamic,' but Eq. (5) uses global scalar weights with no input dependence. It's a static learned filter. If the authors mean 'adaptive' in the sense of trained on data, that's fine, but 'content-aware' is misleading. Fourth, the interpretability analysis—reading the learned weights as neuroscientific evidence—is partially circular, since the weights were optimized to make latents predictable from fMRI, not to explain the brain. It's suggestive, not evidence. Code is not provided, which makes all of this harder to check.\n\nThe idea still has some merit. A frequency-gating pre-filter is cheap and might help in other VDV AE pipelines, and the weight visualization offers a concrete target for future work on non-uniform bands. But as a paper, the empirical support is too thin and the claims too broad. It deserves serious refereeing, because the idea is plausible and the writing is competent, but it needs major revision: factual reframing of what FreqSelect does and doesn't improve, multi-subject results with error bars, and a reining-in of the interpretability claims. If the authors can show even a reliable small gain in SSIM/CLIP across subjects, that's a modest but citable contribution.","headline":"A plausible frequency-gating idea whose central claim of consistent improvement is contradicted by the paper's own Table 1.","tokens_in":13953,"tokens_out":3105,"would_cite":false,"duration_ms":29309,"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":"FreqSelect filters fMRI images by frequency band to boost reconstruction","keywords":["fMRI-to-image reconstruction","spatial frequency","frequency band filtering","natural scenes dataset","very deep variational autoencoder","latent diffusion","brain decoding","frequency gating"],"falsifier":"Run FreqSelect with N=16 on all four subjects (sub1, sub2, sub5, sub7) with multiple random seeds, report per-subject and per-seed metric distributions with error bars, and check whether the SSIM and CLIP gains over Brain-Diffuser replicate and whether the improvement is consistent across more than a single metric subset. A simpler direct test: replace the learned scalar weights with a fixed low-pass filter that passes only the lowest band, and compare its SSIM and CLIP scores; if the fixed filter matches FreqSelect, the learned weights add nothing.","tokens_in":12928,"feed_emoji":"🧠","tokens_out":1496,"duration_ms":16173,"temperature":0.7,"pith_summary":"The paper proposes FreqSelect, a lightweight module that splits an input image into radial spatial-frequency bands and learns a single scalar weight per band before the image is encoded into the latent space of a frozen very deep variational autoencoder. The weights are trained jointly with a ridge regressor that predicts VDV-AE latents from fMRI activity, so the module learns to pass low frequencies that carry global structure and to suppress high frequencies dominated by noise. The authors claim this consistently improves fMRI-to-image reconstruction on the Natural Scenes Dataset compared with the Brain-Diffuser pipeline, most clearly on SSIM and CLIP scores at sixteen bands, and that the learned pass-through rates offer neuro-scientific insight into how visual frequencies are represented in the brain.","feed_headline":"Frequency filtering sharpens fMRI image reconstruction","feed_subtitle":"A lightweight module learns which spatial frequency bands carry neural signal, improving SSIM and CLIP scores on the Natural Scenes Dataset.","key_machinery":"The central mechanism is a radial band-pass decomposition of the 64x64 input image into N concentric frequency bands using binary Fourier-domain masks, followed by Eq. (5), a weighted average of the inverse-DFT band-limited images with sigmoid-activated scalar weights alpha_i = sigma(w_i). The weighted fused image is passed through a frozen VDV-AE encoder, and the weights and a ridge regressor are trained end-to-end to minimize latent-space mean squared error.","core_discovery":"FreqSelect's central claim is that a fixed, learned set of radial frequency-band weights, applied to the stimulus image before VDV-AE encoding, improves fMRI-to-image reconstruction by acting as a content-aware gate that suppresses noise-dominated high frequencies while preserving informative low frequencies. The authors report that at N=16 the method outperforms Brain-Diffuser on SSIM (0.2961 vs 0.293) and CLIP score (92.65% vs 92.5%), while remaining competitive on other metrics. They also claim generalization across subjects and interpretable learned pass-through rates, with the model consistently assigning the highest weight to the lowest frequency band.","pith_inferences":["The paper's own numbers show the 'consistent improvement' claim is overstated: at N=16 FreqSelect beats Brain-Diffuser on only SSIM and CLIP, and loses on PixCorr, AlexNet(2/5), Inception, EfficientNet-B, and SwAV on subject 1.","Because the weights in Eq. (5) are scalars shared across all images and subjects, the 'content-aware' and 'dynamic' description in the abstract does not match the implementation; a per-image or per-subject gate would be needed to test whether adaptivity is what drives the gain.","A testable extension is to fix the frequency partition but replace scalar weights with image-dependent predictions (e.g., a small network on the FFT magnitude), then compare against the fixed-weight version to see whether 'content-aware' gating adds anything beyond a static low-pass filter.","The reported evaluation is on subject 1 only without error bars, so cross-subject generalizability, including the claim of 'generalizes across subjects,' is not empirically established in the quantitative results."],"forward_implications":["If the central claim holds, FreqSelect becomes a cheap plug-in improvement for VDV-AE-diffusion fMRI decoders, requiring only N scalar parameters and no retraining of the generative backbones.","The learned pass-through rates would provide a data-driven map of which spatial-frequency bands carry decodable neural information, potentially connecting model training dynamics to known band-pass tuning in early visual cortex.","The architecture suggests that suppressing high-frequency input components can reduce overfitting to fMRI noise, which may transfer to other neuroimaging modalities with frequency-dependent SNR.","The finding that N=8 hurts performance while N=16 helps indicates that band granularity is a critical design parameter, motivating non-uniform or data-driven frequency partitions.","The reported gains on SSIM and CLIP but losses on several other metrics imply a trade-off between structural fidelity and pixel-level or feature-level accuracy that practitioners would need to weigh."],"supporting_citations":[{"why":"Brain-Diffuser is the baseline pipeline that FreqSelect plugs into; it supplies the VDV-AE configuration, the latent extraction details, and the quantitative and qualitative comparison targets.","marker":"[19]"},{"why":"The Natural Scenes Dataset provides all fMRI data and stimuli; the paper's subject selection, train/test split, and stimulus paradigm come from this dataset.","marker":"[1]"},{"why":"Very Deep VAE is the frozen encoder whose latent space is the supervision target; the hierarchical latent structure in Eq. (6)-(7) is taken directly from this work.","marker":"[5]"},{"why":"AutoKL, part of the latent diffusion model, is used for Stage 2 refinement; the noise-prediction objective in Eq. (9)-(10) relies on this formulation.","marker":"[23]"},{"why":"Versatile Diffusion is the Stage 2 latent diffusion model that refines VDV-AE outputs, supplying the U-Net, cross-attention, and conditioning mechanism.","marker":"[36]"},{"why":"CLIP-ViT/L-14 provides the semantic embeddings used for conditioning and the CLIP score metric; the Stage 2 training uses CLIP-based features.","marker":"[21]"},{"why":"AlexNet feature correlations (layers 2 and 5) serve as mid-level metrics in the evaluation.","marker":"[11]"},{"why":"SSIM is a low-level structural similarity metric used in the comparison tables.","marker":"[34]"}],"fun_headline_variants":["FreqSelect: frequency-aware gating sharpens fMRI-to-image","Frequency filtering sharpens fMRI image reconstruction","Adaptive frequency gating improves fMRI image decoding","FreqSelect teaches networks which spatial frequencies matter","FreqSelect filters frequencies to sharpen fMRI decoding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a single global set of learned frequency-band weights, fixed after training and shared across all images and subjects, is sufficient to realize the claimed content-aware, dynamic gating and to yield consistent improvements across metrics and subjects.","fun_headline_variants_meta":{"raw":{"variants":["FreqSelect: frequency-aware gating sharpens fMRI-to-image","Frequency filtering sharpens fMRI image reconstruction","Adaptive frequency gating improves fMRI image decoding","FreqSelect teaches networks which spatial frequencies matter","FreqSelect filters frequencies to sharpen fMRI decoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000767,"raw_usage":{"total_tokens":3381,"prompt_tokens":904,"completion_tokens":2477,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":2404}},"tokens_in":520,"tokens_out":2477,"duration_ms":20592,"temperature":1.0,"reasoning_tokens":2404,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:31:37.216288+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FreqSelect with N=16 on all four subjects (sub1, sub2, sub5, sub7) with multiple random seeds, report per-subject and per-seed metric distributions with error bars, and check whether the SSIM and CLIP gains over Brain-Diffuser replicate and whether the improvement is consistent across more than a single metric subset. A simpler direct test: replace the learned scalar weights with a fixed low-pass filter that passes only the lowest band, and compare its SSIM and CLIP scores; if the fixed filter matches FreqSelect, the learned weights add nothing.","supporting_citations":[{"cited_title":"Allen, Ghislain St-Yves, Yihan Wu, Jesse L","cited_arxiv_id":null,"evidence_quote":"The Natural Scenes Dataset provides all fMRI data and stimuli; the paper's subject selection, train/test split, and stimulus paradigm come from this dataset."},{"cited_title":"Very deep vaes generalize autoregressive models and can outperform them on images","cited_arxiv_id":null,"evidence_quote":"Very Deep VAE is the frozen encoder whose latent space is the supervision target; the hierarchical latent structure in Eq. (6)-(7) is taken directly from this work."},{"cited_title":"Versa- tile diffusion: Text, images and variations all in one diffusion model","cited_arxiv_id":null,"evidence_quote":"Versatile Diffusion is the Stage 2 latent diffusion model that refines VDV-AE outputs, supplying the U-Net, cross-attention, and conditioning mechanism."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"AlexNet feature correlations (layers 2 and 5) serve as mid-level metrics in the evaluation."}],"review_version":1}