{"id":"d526c5e3-c0a4-4eee-a1fc-f5f803398983","arxiv_id":"2607.08033","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":9,"one_line_summary":"A sequence-dominated Mamba architecture with Retinex priors achieves fast, unsupervised low-light enhancement of spacecraft imagery on resource-limited hardware.","lead":"The paper builds SCI-Mamba, a lightweight neural network that enhances dark images of spacecraft for on-orbit servicing. It matters because it runs fast on limited satellite hardware while suppressing background noise better than prior methods.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Table 3's no-reference metrics show the raw unenhanced input scores better than SCI-Mamba on all three metrics (NIQE 20.21<20.39, BRISQUE 54.98<56.71, PIQE 54.17<54.62), undermining the quantitative basis for the visual quality claim.","rationale":"The reader correctly identified two gaps (no-reference metrics may not transfer; no downstream task evaluation) and correctly flagged the illumination monitor's hyperparameter sensitivity. However, the reader did not fully appreciate the severity of the metrics issue: the raw input outperforms SCI-Mamba on all three no-reference metrics, meaning the quantitative evidence as presented actually contradicts the visual quality improvement claim rather than merely being questionable. The reader's illumination monitor concern is valid but less load-bearing — it is a training-time loss component, and even imperfect masks would degrade rather than collapse performance given the other four loss terms. The efficiency claims (23.05% speedup, order-of-magnitude memory reduction) are well-supported by the ablation in Table 2 and stand on solid ground. The architectural innovation is real and the code is available. But the visual quality claim — half of the central 'trade-off' — lacks quantitative support. The verdict should remain CONDITIONAL: the engineering contribution is sound, but the paper needs downstream task evaluation or full-reference metrics to substantiate the quality improvement claim. The reader's recommendation for downstream task evaluation is exactly right and would settle this concern.","tokens_in":22700,"tokens_out":2388,"duration_ms":90858,"concrete_test":"Run a pre-trained spacecraft pose estimation or component detection model (e.g., from SPEED+ benchmark) on both raw low-light images and SCI-Mamba enhanced images from the Space Dark-1.0 test set. Compare task-level metrics (pose error, detection mAP). If enhanced images do not yield measurably better downstream performance than raw inputs, the practical enhancement claim is unsupported regardless of subjective visual appearance.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is a 'favorable trade-off between visual quality, noise suppression, and hardware efficiency.' The efficiency side is well-supported by Table 2. However, the visual quality side rests on Table 3's no-reference metrics and subjective qualitative comparisons. Examining Table 3 carefully: lower values indicate better quality for all three metrics. The raw input achieves NIQE=20.21, BRISQUE=54.98, PIQE=54.17. SCI-Mamba achieves NIQE=20.39, BRISQUE=56.71, PIQE=54.62. The raw unenhanced input is better than SCI-Mamba on ALL THREE metrics. The paper acknowledges this for PIQE ('marginally higher than the unprocessed input') but obscures it for NIQE by comparing only against competing methods (claiming 'second only to LLFormer') while omitting that the raw input outperforms both. For BRISQUE, the paper frames 56.71 as a 'practical balance' without noting the raw input's 54.98 is lower. This means the quantitative evidence does not support improved visual quality — it shows the opposite. The entire visual quality claim then depends solely on subjective visual comparisons in Figures 8-13, which are not independently verifiable. Without either full-reference metrics (requiring paired ground truth) or downstream task evaluation (e.g., pose estimation accuracy with vs. without enhancement), the claim that SCI-Mamba improves visual quality for operational use is not quantitatively substantiated. This is more load-bearing than the illumination monitor concern because it strikes at the empirical evidence for the central claim itself, not just one architectural component.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper proposes SCI-Mamba, an unsupervised low-light image enhancement network designed for non-cooperative spacecraft imagery. The architecture combines the self-calibrated illumination learning of SCI++ with a 1D-only sequence processing pipeline built on VMamba's state-space modeling, performing a single 2D-to-1D flattening at input and a single 1D-to-2D reconstruction at output to eliminate layer-wise cross-dimensional overhead. A multi-objective loss includes a prior-guided illumination monitor that auto-generates target illumination maps from input brightness and texture features to drive differentiated foreground/background enhancement. The authors also construct Space Dark-1.0, a multi-source dataset combining synthetic renders, darkroom hardware-in-the-loop captures, and real on-orbit footage. Efficiency results (Table 2) demonstrate strong speed and memory advantages over Transformer and Mamba competitors, with a 23.05% speedup over the 2D-scan ablation baseline.","tokens_in":23678,"tokens_out":1161,"duration_ms":113457,"significance":"The paper addresses a genuine engineering gap: existing low-light enhancement methods are either too computationally heavy (Transformer/Mamba variants with quadratic or layer-wise overhead) or lack global context modeling (CNN methods). The one-time bidirectional conversion design is a concrete, verifiable architectural contribution, and the ablation in Table 2 (SCI-Mamba vs. SCI-Mamba(2D)) directly supports the 23.05% speedup claim. The release of Space Dark-1.0, integrating real ADRAS-J orbital footage with hardware-in-the-loop data, is a valuable resource contribution. Code availability is stated. The work targets a well-motivated application domain with strict hardware constraints.","major_comments":[{"comment":"Table 3 and the visual quality claim. The paper's central claim is a 'favorable trade-off between visual quality, noise suppression, and hardware efficiency.' The efficiency side is well-supported. However, on all three no-reference metrics in Table 3, the raw unenhanced input scores better than SCI-Mamba (NIQE: 20.21 vs. 20.39; BRISQUE: 54.98 vs. 56.71; PIQE: 54.17 vs. 54.62). The text acknowledges this for PIQE but obscures it for NIQE and BRISQUE by framing SCI-Mamba's scores favorably only relative to competing enhancement methods. This means the quantitative evidence does not support improved visual quality over the raw input. The visual quality claim then rests entirely on subjective Figures 8-13. The authors should either (a) add full-reference metrics on the synthetic subset where ground truth is available, or (b) add a downstream task evaluation (e.g., pose estimation or targetd","section":null}],"minor_comments":[{"comment":"Table 1 caption says 'Training Overhead Comparison Across Unsupervised Enhancement Algorithms' but includes no Transformer or Mamba methods, only CNN-based unsupervised competitors. The caption should clarify this scope or the table title should be adjusted to avoid confusion with the inference comparison in Table 2.","section":null},{"comment":"Section 4.2: The training set combines 8000 samples from Space Dark-1.0, 1600 from ExDark, and 2400 from SPEED+. The test set is described as 1400 unseen samples from Space Dark-1.0. It is unclear whether the 8000 training images and 1400 test images are drawn from the same 9400-image pool or from disjoint partitions. Please clarify the exact train/test split.","section":null},{"comment":"The large number of hand-tuned hyperparameters (alpha_1 through alpha_5, beta weights, lambda_d, lambda_b, sigma values, percentile thresholds, etc.) makes reproducibility challenging. A sensitivity analysis on at least the most impactful parameters (e.g., alpha_5 for Loss5, lambda_d/lambda_b for the illumination monitor) would help readers understand robustness.","section":null},{"comment":"Figures 8-13: The qualitative comparisons would be more informative if zoomed-in crops of specific regions (e.g., solar panel edges, background noise patches) were provided alongside full images, as is standard in image restoration papers. At current scale, fine differences are difficult to verify for readers.","section":null},{"comment":"Reference [27] (SISIFOS) and several others cite 2026 conference/forum dates. If these are accepted/published, the citations should be updated to final DOIs/venue information rather than pending.","section":null},{"comment":"Eq. (1): The notation I_3^{(-1)} is described as 'element-wise Hadamard inverse' but the superscript notation could be confused with matrix inversion. Consider using I_3^{odot (-1)} or explicitly stating 'element-wise reciprocal' to match standard conventions.","section":null}],"recommendation":"major_revision","confidential_remarks":"The core architectural efficiency contribution is solid and well-ablated. The main concern is that the visual quality claim is not quantitatively supported — the no-reference metrics actually favor the raw input. This is a known limitation of NR metrics on low-light images (dark images can score artificially well), but the authors need to address it with either full-reference metrics on synthetic data or downstream task evaluation. Without this, the paper's trade-off claim is only half-substantiated. The prior-guided illumination monitor's circularity (training to match auto-generated heuristic masks) is a secondary concern; it is a design choice that could work if the heuristics are reasonable, but the paper would benefit from an ablation showing the impact of Loss5."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The stress-test concern lands hard. Look at Table 3: lower is better for all three metrics, and the raw unenhanced input beats SCI-Mamba across the board (NIQE 20.21 vs 20.39, BRISQUE 54.98 vs 56.71, PIQE 54.17 vs 54.62). The paper acknowledges this for PIQE but sidesteps it for NIQE and BRISQUE. So the quantitative evidence doesn't support improved visual quality — it shows the opposite. The visual quality claim then rests entirely on subjective figures, which aren't independently verifiable. That's a load-bearing problem for the paper's central trade-off argument. The efficiency side (Table 2) is solid and well-ablated, but the quality side needs full-reference metrics or downstream task evaluation to be convincing. The prior-guided illumination monitor (Loss5) is a secondary concern — ~20 hand-tuned hyperparameters with no sensitivity analysis is not ideal, but it's a training signal, not the core claim. The circularity is mild. What's genuinely new: the fully 1D sequence-dominated pipeline that flattens once at input and reconstructs once at output, avoiding layer-wise 2D/1D conversion. The 23% speedup over the 2D-scan ablation is real. Space Dark-1.0 is a useful dataset contribution. Code is available. This is an engineering paper for a niche but important domain, and the architecture is legitimate. But the evaluation gap is too large to ignore — no-reference metrics designed for natural ground scenes may not transfer to orbital imagery, and without downstream task metrics (pose estimation, detection), the operational utility is asserted, not demonstrated. The paper needs either full-reference metrics on paired data or downstream task results to substantiate the quality claim. Worth a serious referee because the architecture and dataset are real contributions, but the evaluation must be addressed.","headline":"The 1D-only Mamba pipeline is a real engineering contribution, but Table 3 undercuts the visual quality claim — the raw input scores better than SCI-Mamba on all three no-reference metrics.","tokens_in":23672,"tokens_out":493,"would_cite":false,"duration_ms":77511,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Flatten once, enhance in 1D: a lightweight pipeline for dark spacecraft imagery","keywords":[],"falsifier":"A test set of real orbital images where the spacecraft target is so small or so dim that the Otsu-based bright-region map and gradient-based texture map fail to produce a coherent structural mask. If the auto-generated target illumination map misidentifies background noise as texture or misses the actual spacecraft structure, the differentiated enhancement would either amplify background noise or leave the target under-enhanced, showing no advantage over the global brightness amplification methods the paper claims to surpass.","tokens_in":22984,"feed_emoji":"🛰️","tokens_out":995,"duration_ms":150851,"temperature":0.7,"pith_summary":"The paper proposes SCI-Mamba, an unsupervised low-light image enhancement network designed for non-cooperative spacecraft observations. The core architectural claim is that by flattening a 2D image into a 1D sequence exactly once at the input and performing all illumination estimation and self-calibration operations purely within that 1D sequence domain (using a redesigned visual state-space module), one can eliminate the redundant layer-wise 2D-to-1D and back conversions that plague existing Mamba and Transformer restoration networks. This yields a 23% inference speedup over the 2D-scan baseline and an order-of-magnitude memory reduction relative to competing Mamba methods, making the model feasible for resource-constrained spaceborne hardware. A second key claim is that a prior-guided illumination monitor—built entirely from the input image's own brightness and gradient statistics—can automatically generate a target illumination map that drives the network to enhance texture-rich spacecraft regions while suppressing noise in flat deep-space backgrounds, achieving differentiated enhancement without paired training data.","feed_headline":"Flatten once, enhance in 1D: a lightweight pipeline for dark spacecraft imagery","feed_subtitle":"By keeping all illumination estimation in 1D sequence space and auto-generating foreground masks from input gradients, the method runs 4-18x","key_machinery":"The VSS1D module: a residual block that receives a globally flattened 1D patch sequence and uses an embedded SCSG (Sequential Cross-scan Generator) to produce four directional scan sequences (row-forward, column-forward, row-reverse, column-reverse) purely via 1D permutation and reversal operations. Each sequence is processed by Mamba's selective state-space model, then an SCSR (Sequential Cross-scan Restorer) reverses the permutations and fuses all four streams back to row-major order inside 1D space. This replaces VMamba's standard SS2D, which requires 2D feature map reconstruction at every layer.","core_discovery":"The paper's central discovery is that a fully sequence-dominated pipeline, which performs a single global 2D-to-1D flattening at input and a single 1D-to-2D reconstruction at output, can serve as a complete and efficient substitute for the repeated cross-dimensional transformations used in standard visual Mamba architectures for low-light enhancement. Within this 1D-only domain, a Sequential Cross-scan Generator (SCSG) rearranges the base sequence into four directional traversal streams to recover the spatial context lost during flattening, and a Sequential Cross-scan Restorer (SCSR) fuses them back—all without ever reconstructing a 2D feature map. Combined with a self-calibrated Retinex框架, ","pith_inferences":[],"forward_implications":["If the one-flatten design principle generalizes, other image-to-sequence tasks (e.g., denoising, super-resolution) currently using VMamba-style architectures could adopt the VSS1D module to reduce memory and latency without sacrificing spatial modeling.","The prior-guided illumination monitor, if robust, suggests that unsupervised differentiated enhancement can be achieved for any domain with a strong foreground-background radiometric separation (e.g., underwater imagery, medical imaging) by auto-generating target illumination maps from input gradients and brightness alone.","The Space Dark-1.0 dataset, combining synthetic renders, hardware-in-the-loop darkroom captures, and real orbital footage, provides a benchmark that could standardize evaluation for future space-target perception algorithms under low-light conditions.","If the 1D-sequence Retinex operations (element-wise Hadamard multiplication/division on flattened sequences) prove numerically stable across deeper pipelines, they could replace image-domain Retinex in other multi-stage enhancement frameworks to reduce cross-domain switching overhead."],"fun_headline_variants":["One flatten, one rebuild: 1D-only Mamba pipeline for dark spacecraft images","Skip the 2D round-trips: single-flatten Mamba for low-light space imagery","Flatten once, stay 1D: cross-scan streams recover spatial context without 2D maps","No intermediate 2D maps: 1D-dominated Retinex-Mamba for dark spacecraft images","Four directional 1D scans replace 2D feature maps in space image enhancer"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The prior-guided illumination monitor assumes that Otsu thresholding and morphological operations applied to extremely dark, noisy input images can reliably segment spacecraft foreground from deep-space background. If these heuristic masks fail on degenerate inputs, the differentiated enhancement strategy would collapse into the same indiscriminate noise amplification the paper criticizes in CNN baselines.","fun_headline_variants_meta":{"raw":{"variants":["One flatten, one rebuild: 1D-only Mamba pipeline for dark spacecraft images","Skip the 2D round-trips: single-flatten Mamba for low-light space imagery","Flatten once, stay 1D: cross-scan streams recover spatial context without 2D maps","No intermediate 2D maps: 1D-dominated Retinex-Mamba for dark spacecraft images","Four directional 1D scans replace 2D feature maps in space image enhancer","Self-calibrated Retinex meets 1D-only Mamba for unsupervised dark spacecraft","Stay 1D end to end: cross-scan generator and restorer replace 2D feature maps","Unsupervised Retinex-Mamba enhances dark spacecraft images without 2D detours","No 2D detours: sequential cross-scan handles low-light enhancement entirely in 1D","One flatten, one rebuild: 1D Mamba enhancement runs 4-18x faster for spacecraft"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":1689,"prompt_tokens":569,"completion_tokens":1120,"prompt_tokens_details":null},"tokens_in":569,"tokens_out":1120,"duration_ms":19419,"temperature":1.0,"reasoning_tokens":816,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T01:16:29.208881+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"A test set of real orbital images where the spacecraft target is so small or so dim that the Otsu-based bright-region map and gradient-based texture map fail to produce a coherent structural mask. If the auto-generated target illumination map misidentifies background noise as texture or misses the actual spacecraft structure, the differentiated enhancement would either amplify background noise or leave the target under-enhanced, showing no advantage over the global brightness amplification methods the paper claims to surpass.","supporting_citations":[],"review_version":1}