{"id":"b56d6f61-7bd7-40fa-b709-8c401bf3e465","arxiv_id":"2501.04477","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SpikeCLIP reconstructs low-light spike-camera images by aligning outputs with CLIP class labels and prompts learned from NIQE-selected high-quality images, reporting state-of-the-art no-reference quality scores on U-CALTECH and U-CIFAR.","lead":"SpikeCLIP trains a spike-camera image reconstruction network using CLIP text labels and high-quality reference images instead of paired ground truth, targeting low-light scenes. It reports large gains over prior methods on real-world low-light spike datasets, but its evaluation relies on no-reference metrics that are partly entangled with how the training data was selected.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quantitative superiority claim is at risk of circularity: HQ images are selected by the same NIQE metric used for evaluation, so reported gains may reflect metric optimization rather than genuine perceptual quality.","rationale":"The reader's weakest assumption identified the same load-bearing concern: the HQ dataset is built by minimizing NIQE (Eq. 6), and the paper's primary results are NIQE scores (Table 1). This creates a direct pathway for the training process to optimize the evaluation metric, even though the HQ images come from different scenes than the test set. The prompt loss and class loss are trained in CLIP feature space, which is not a direct NIQE gradient, but the selection of HQ images by NIQE biases the learned prompt toward whatever visual characteristics yield low NIQE in those scenes. Because U-CALTECH and U-CIFAR share the same spike camera and low-light conditions, those characteristics transfer, explaining the unusually large metric gains. This concern is load-bearing because it is the only quantitative evidence for the central claim; without it, the claim rests on a handful of selected visual examples. The proposed test directly manipulates the selection mechanism: if random selection removes the NIQE advantage, the argument is circular. If random selection preserves the advantage, the model may genuinely be improving perceptual quality through the CLIP prompt, and the concern would be resolved. I therefore keep the reader's CONDITIONAL verdict: the paper is plausible and technically coherent, but the central quantitative claim requires an independent evaluation protocol, either a human perceptual study or the selection-ablation test described above. I agree with the reader that the class-label supervision asymmetry is a limitation, but it is secondary to the metric-bias issue and does not change the verdict.","tokens_in":9695,"tokens_out":5452,"duration_ms":54049,"concrete_test":"Rebuild the HQ dataset by replacing Eq. (6)'s argmin NIQE selection with a random choice among the same candidate methods for each real-world scene, keeping the synthetic Spike-REDS portion unchanged. Retrain the full three-stage pipeline with identical hyperparameters and report NIQE, BRISQUE, and PIQE on U-CALTECH and U-CIFAR. If the NIQE lead over WGSE shown in Table 1 narrows to within 10% or reverses, the original gains are largely an artifact of optimizing the evaluation statistic during dataset construction; if the lead persists, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that SpikeCLIP surpasses prior methods on real-world low-light spike datasets rests on Table 1, which reports no-reference metrics NIQE, BRISQUE, and PIQE. But the training signal is not independent of these metrics. In Eq. (6), the HQ dataset is constructed by selecting, for each real-world scene, the reconstruction among TFP, TFI, Spk2ImgNet, SSML, and WGSE that achieves the lowest NIQE. The prompt-loss stage then trains the lightweight network to align its output with the CLIP embedding of this NIQE-selected HQ set, and the fine-tuning stage is supervised by that prompt plus class labels. On the U-CALTECH and U-CIFAR test sets, the same sensor and scene type are used, so the model has effectively been optimized, through a learned CLIP-space proxy, toward the very statistic that is later reported as the headline improvement. Consequently, the large NIQE/BRISQUE/PIQE gains do not by themselves establish superior texture detail or luminance balance; they may instead indicate that the model has learned to exploit the specific bias of these no-reference metrics. A secondary but related issue is that SpikeCLIP uses class labels during training while the compared baselines do not, making the comparison unequal; however, the metric-selection circularity is the more fundamental threat because it undermines the validity of the quantitative evidence itself. The visual examples in Fig. 5 are selected demonstrations and cannot independently resolve this.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SpikeCLIP, a three-stage spike-to-image reconstruction framework for real-world low-light spike streams. Stage 1 trains a lightweight reconstruction network to reproduce TFI results; Stage 2 constructs a high-quality (HQ) image set from synthetic and real datasets, selecting candidate reconstructions with the lowest NIQE score (Eq. 6), and learns prompt embeddings that separate HQ from low-quality images in CLIP space; Stage 3 fine-tunes the network with a class-label InfoNCE loss (Eq. 10) and a prompt loss (Eq. 9). Experiments on U-CALTECH and U-CIFAR report large improvements over prior methods in NIQE, BRISQUE, and PIQE, higher CLIP classification accuracy, and very low parameter count, FLOPs, and latency. The central claim is that this CLIP-based supervision, using class labels and unpaired HQ images, suffices to train a lightweight network that recovers texture-rich and brightness-balanced images under low light without paired ground truth.","tokens_in":9982,"tokens_out":3048,"duration_ms":32411,"significance":"If the quantitative claims hold, the framework is a meaningful contribution: it removes the requirement of paired spike-sharp supervision in real-world low-light spike reconstruction, exploits CLIP's text-image alignment as a supervisory signal, and demonstrates that a lightweight network can suffice when the supervision is semantic. The paper also releases code, which supports reproducibility. However, the evaluation protocol is too weak to support the central claim as it stands: the training-target selection and the headline evaluation metric are not independent, and the method uses class labels that the baselines do not. The core idea is promising, but the evidence for superior reconstruction quality requires independent validation.","major_comments":[{"comment":"The evaluation of reconstruction quality is circular with respect to the HQ dataset construction. In Eq. (6), the HQ images used as positive training targets are selected by minimizing NIQE among candidate reconstructions on real-world datasets, and the main results in Table 1 are reported as NIQE improvements. Because the prompt-loss and fine-tuning stages align the LRN output with embeddings of these NIQE-selected images, a large part of the reported NIQE gain is expected by construction rather than being evidence of genuine perceptual improvement. The claim that SpikeCLIP 'significantly enhances texture details and luminance balance' is therefore not established by Table 1. Please provide evaluation with a metric not involved in dataset construction, such as a human perceptual study, or report scores on an HQ set selected without NIQE.","section":"Prompt Learning, Eq. (6) and Experimental Results, Table 1"},{"comment":"The comparison with prior methods is unequal because SpikeCLIP is trained with ground-truth class labels for the reconstructed scene, whereas TFP, TFI, TFSTP, SSML, SpikeFormer, RSIR, Spk2ImgNet, and WGSE do not use label information. The gains in Table 1 may reflect the additional semantic supervision rather than a fundamentally better spike-to-image mapping. In particular, the classification accuracy reported in Table 2 measures alignment with the text labels, not image reconstruction quality. Please compare against baselines that also receive the same class-label information, or explicitly justify why the comparison remains fair for the reconstruction-quality claim.","section":"Fine Reconstruction, Eq. (10), and Table 1"},{"comment":"All quantitative quality evidence rests on no-reference metrics NIQE, BRISQUE, and PIQE, which are known to correlate imperfectly with perceptual quality and can be optimized artificially. There is no real-world paired ground truth, and Fig. 5 shows only selected qualitative examples. Please report variability across the 1,000 test samples, include statistical significance tests and error bars, and provide a more complete qualitative panel rather than a single selected frame per scene. This would substantially increase confidence that the improvements are perceptual and not metric-specific.","section":"Experimental Results, Table 1 and Fig. 5"}],"minor_comments":[{"comment":"The caption contains a typo: 'U-CALTEHC' should be 'U-CALTECH'.","section":"Fig. 5 caption"},{"comment":"The prompt loss in Eq. (9) is the negative softmax probability without a logarithm; please clarify whether this is intentional or whether a cross-entropy formulation was intended, since the loss magnitude and gradient scale differ.","section":"Eq. (9)"},{"comment":"The columns 'Params' and 'Flops' in Table 3 appear to omit units; they are presumably millions and giga-FLOPs, respectively, and should be labeled consistently with Table 1.","section":"Table 3"},{"comment":"The text says the image with the 'best non-reference metric NIQE score' is selected, but because lower NIQE is better, 'best' should be 'lowest' for clarity.","section":"Eq. (6)"},{"comment":"The text says bold and underline distinguish best and second-best results, but Table 1 shows only bold entries and no underlined entries; please align the formatting with the stated policy.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a computer-vision venue and the core idea is worth pursuing. The main risk is the circular use of NIQE in both dataset construction and evaluation; if the authors add an independent evaluation protocol and address the unequal supervision with class labels, the contribution could be publishable. I recommend major revision rather than rejection because the flaws are in the evaluation design and can plausibly be fixed within the paper's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing you should know: SpikeCLIP is a genuinely new training paradigm for spike-to-image reconstruction—CLIP text supervision, learnable prompts, and unpaired high-quality images instead of paired ground truth—but the headline numbers are not trustworthy because the same no-reference metric, NIQE, is used both to build the training set and to evaluate the results.\n\nWhat is actually new and good: applying CLIP-LIT/CoOp-style prompt learning to spike cameras is a reasonable transfer, and the three-stage pipeline (coarse reconstruction, prompt learning, fine reconstruction) is clearly described and sensible. The paper also shows that a tiny network (0.186M params) can produce plausible reconstructions when guided by semantic supervision, which is a useful data point for the subfield. The code is available, and the ablations are consistent. I believe the authors know what they are doing methodologically.\n\nWhere it gets soft: Equation (6) constructs the HQ dataset by selecting, for each real-world scene, the reconstruction with the lowest NIQE among TFP, TFI, Spk2ImgNet, SSML, and WGSE. Table 1 then reports NIQE as the primary quality metric. So the model is effectively trained to match images that were chosen to minimize NIQE, and then judged with NIQE. That is a real circularity, and the stress-test note is right to flag it. The large NIQE/BRISQUE/PIQE gains do not establish genuine perceptual superiority; they may partly reflect optimization toward those specific statistics. A second, related problem is that the model sees class labels during training while the baselines do not, making the comparison unequal. That alone would not sink the paper, but combined with the metric loop it means the central claim of \"surpasses previous methods\" is not convincingly supported.\n\nMinor issues: the reliance on class labels is an unstated limitation—if you don't have labels for the scene, this training paradigm doesn't apply. The visual examples in Fig. 5 are selected and can't carry the weight. No human study, no paired data in realistic conditions, and no held-out metric test.\n\nThe core idea deserves attention from the neuromorphic vision community. It is a plausible direction, and the paper is worth reading for the training paradigm alone. But the evaluation needs a major rework before the quantitative claims can be taken at face value. A serious referee should ask for an independent evaluation: a human perceptual study, validation on paired synthetic low-light data, or at least a protocol where the training-target metric is not the same as the evaluation metric.\n\nMy recommendation: send it to peer review, not desk reject. The novelty and the code are real, and the circularity is fixable. But the paper should come back with substantially strengthened evidence.","headline":"Novel CLIP-based training paradigm for spike reconstruction, undermined by an evaluation loop where NIQE selects the training targets and then scores the results.","tokens_in":10513,"tokens_out":1632,"would_cite":false,"duration_ms":18665,"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":"SpikeCLIP trains low-light spike-to-image reconstruction from CLIP text and unpaired high-quality images, and reports better reconstruction quality on U-CALTECH and U-CIFAR than prior methods.","keywords":["spike camera","image reconstruction","low-light imaging","CLIP","prompt learning","contrastive learning","neuromorphic vision","no-reference image quality"],"falsifier":"Take a held-out set of real low-light spike streams not used in training, reconstruct with SpikeCLIP and with WGSE or TFI, and run a forced-choice human perceptual study with equalized display. If human raters prefer SpikeCLIP no more often than chance, or if a perceptual metric that was not used to build the HQ dataset (for example, a trained opinion-aware image quality model) shows no advantage, the claimed perceptual improvement would not hold.","tokens_in":9485,"feed_emoji":"📷","tokens_out":6818,"duration_ms":59308,"temperature":0.7,"pith_summary":"SpikeCLIP aims to show that a spike-to-image network can be trained for low-light scenes without any paired sharp ground truth, using only the class label of the scene and an unpaired collection of high-quality images. The paper argues that CLIP's shared text-image space turns these weak signals into a strong training signal, so a lightweight 0.186M-parameter network can reconstruct texture-rich, brightness-balanced images. This matters because previous supervised and self-supervised methods degrade on real low-light spikes, where synthetic training pairs do not match the noise and pseudo-labels lose quality. If the claim holds, class labels plus unpaired high-quality photos are sufficient supervision for practical spike-camera reconstruction under extreme lighting.","feed_headline":"CLIP text supervision sharpens low-light spike camera images","feed_subtitle":"A tiny network beats prior spike-reconstruction methods using only labels and unpaired photos.","key_machinery":"The load-bearing mechanism is CLIP as a joint text-image embedding space, used as the only perceptual supervisor. Two learnable prompts, initialized in the style of CoOp, are optimized to represent the distributions of high-quality and low-quality reconstructions; the reconstruction network is then trained with a decoupled class loss and prompt loss in CLIP's shared embedding space. A supporting mechanism is the HQ-image pipeline, which merges outputs of several reconstruction algorithms (TFP, TFI, Spk2ImgNet, SSML, WGSE) and keeps the lowest-NIQE image per sample, while a TFI-based coarse reconstruction gives the lightweight network a stable initialization.","core_discovery":"On its own terms, the paper's central claim is that SpikeCLIP surpasses previous spike-to-image reconstruction methods on the real-world low-light datasets U-CALTECH and U-CIFAR, and that it does so by replacing ground-truth supervision with two CLIP-based losses. The method has three stages: a coarse stage trains a lightweight reconstruction network to output the model-based TFI estimate; a prompt-learning stage builds a high-quality image dataset from synthetic and real reconstructions and learns CLIP prompts that separate high-quality from low-quality images; and a fine stage optimizes the network with a class loss (InfoNCE against the class label's text feature) plus a prompt loss that pulls reconstructed images toward the high-quality prompt. The paper reports NIQE reductions of 37.9% on U-CALTECH and 43.1% on U-CIFAR over the best compared method, along with lower BRISQUE and PIQE scores, a 64% classification accuracy on U-CALTECH, and a 0.186M-parameter network with 0.540 ms latency.","pith_inferences":["Editorial inference: because the supervision is purely semantic, the same recipe should transfer to other degradation types such as motion blur, haze, or extreme noise wherever a CLIP-recognizable label exists; a direct test would be applying SpikeCLIP to normal-light or deblurring spike datasets.","Editorial inference: the HQ dataset is built by minimizing NIQE, and the headline metrics are also NIQE-family, so a cleaner test of perceptual quality would use a held-out human rating study or a metric not involved in dataset construction.","Editorial inference: the reported classification accuracy suggests the reconstruction network is learning semantic features from spikes even before image quality is perfect, hinting that joint reconstruction-recognition systems could be trained end-to-end with the same losses, a direction the paper does not develop."],"forward_implications":["If the claim holds, low-light spike-to-image reconstruction no longer requires paired sharp ground truth; class labels plus unpaired high-quality images suffice, removing the synthetic-to-real domain gap as a training obstacle.","A 0.186M-parameter network with 0.540 ms latency is enough to beat much larger supervised models, suggesting CLIP supervision can make low-power spike cameras practical for edge deployment.","Because outputs are aligned with CLIP's text-image features, reconstructed images can be directly classified or matched to text without a separate recognition head.","On U-CALTECH and U-CIFAR, the method reports the best NIQE, BRISQUE, and PIQE among the compared methods, with the reported margins shown in the results table."],"supporting_citations":[{"why":"Supplies the CLIP model whose text-image embedding space carries all of SpikeCLIP's supervision signals.","marker":"Radford et al. 2021"},{"why":"Provides the CLIP-LIT prompt-learning framework that the paper adapts for high-quality versus low-quality prompt separation.","marker":"Liang et al. 2023"},{"why":"Provides CoOp, the learnable prompt optimization approach used to obtain the HQ and LQ prompt representations.","marker":"Zhou et al. 2022"},{"why":"Defines TFP and TFI, with TFI used as the coarse reconstruction target and as one of the candidate reconstructions in the HQ image pipeline.","marker":"Zhu et al. 2019"},{"why":"Provides the SSML self-supervised baseline and one of the reconstruction methods whose outputs feed the HQ image generation pipeline.","marker":"Chen et al. 2022"},{"why":"Supplies the UHSR dataset containing the U-CALTECH and U-CIFAR real-world low-light spike streams and class labels used for training and evaluation.","marker":"Zhao et al. 2024a"},{"why":"Defines the NIQE no-reference quality metric used both to select images for the HQ dataset and to evaluate reconstruction quality.","marker":"Mittal, Soundararajan, and Bovik 2012"},{"why":"Provides the WGSE supervised reconstruction method used to generate HQ images on synthetic data and serves as the strongest prior baseline.","marker":"Zhang et al. 2023"},{"why":"Provides the Spk2ImgNet supervised baseline and one of the candidate algorithms in the HQ image generation pipeline.","marker":"Zhao et al. 2021"},{"why":"Defines the InfoNCE contrastive loss that the class loss adapts for aligning reconstructed images with class-label text features.","marker":"Oord, Li, and Vinyals 2018"}],"fun_headline_variants":["SpikeCLIP uses text prompts to sharpen low-light spike camera images","CLIP supervision beats prior methods on low-light spike reconstruction","Text-description supervision improves spike image brightness and detail","Tiny 0.186M network with CLIP losses excels at low-light spike video","Spike-CLIP: text-guided reconstruction for clearer low-light images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that NIQE and the other no-reference metrics measure true reconstruction quality, even though the training's high-quality dataset was built by selecting the lowest-NIQE image, so the reported gains may partly reflect optimizing that same statistic.","fun_headline_variants_meta":{"raw":{"variants":["SpikeCLIP uses text prompts to sharpen low-light spike camera images","CLIP supervision beats prior methods on low-light spike reconstruction","Text-description supervision improves spike image brightness and detail","Tiny 0.186M network with CLIP losses excels at low-light spike video","Spike-CLIP: text-guided reconstruction for clearer low-light images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000676,"raw_usage":{"total_tokens":3091,"prompt_tokens":976,"completion_tokens":2115,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":2022}},"tokens_in":592,"tokens_out":2115,"duration_ms":13986,"temperature":1.0,"reasoning_tokens":2022,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:31:10.507354+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out set of real low-light spike streams not used in training, reconstruct with SpikeCLIP and with WGSE or TFI, and run a forced-choice human perceptual study with equalized display. If human raters prefer SpikeCLIP no more often than chance, or if a perceptual metric that was not used to build the HQ dataset (for example, a trained opinion-aware image quality model) shows no advantage, the claimed perceptual improvement would not hold.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CLIP-LIT prompt-learning framework that the paper adapts for high-quality versus low-quality prompt separation."},{"cited_title":"C.; and Liu, Z","cited_arxiv_id":null,"evidence_quote":"Provides CoOp, the learnable prompt optimization approach used to obtain the HQ and LQ prompt representations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines TFP and TFI, with TFI used as the coarse reconstruction target and as one of the candidate reconstructions in the HQ image pipeline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SSML self-supervised baseline and one of the reconstruction methods whose outputs feed the HQ image generation pipeline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the WGSE supervised reconstruction method used to generate HQ images on synthetic data and serves as the strongest prior baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Spk2ImgNet supervised baseline and one of the candidate algorithms in the HQ image generation pipeline."}],"review_version":1}