{"id":"ef35f77d-3e79-47ae-9c9d-5f393fbb5e9a","arxiv_id":"2509.05000","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A VLM-prompt-guided network that jointly does wavelet-domain degradation suppression and spatial-domain fusion to merge degraded infrared and visible images end to end.","lead":"This paper presents an image fusion system that merges infrared and visible camera images even when both are degraded, using text-based prompts that tell the network what kind of damage to look for. It matters because it could remove the fragile pre-processing steps current systems need, helping night-time and bad-weather perception in driving and surveillance.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table III's 'consistently outperforms' rests solely on four sharpness/contrast metrics that can reward over-enhancement; the paper's own training loss uses degradation-free references, so the omitted reference-based check is decisive.","rationale":"The reader's weakest assumption and mine coincide: the no-reference metrics in Table III do not establish the headline. The internal inconsistency with the training loss makes the omission checkable rather than a matter of taste. Training against I_ref_ir/I_ref_vi proves clean references exist on at least training images; the EMS test split is drawn from the same dataset, so it is plausible they exist there too. If they do, computing reference-based fidelity is straightforward and decisive. If reference metrics reverse or shrink the ranking, the abstract's 'superior fusion performance' is overclaimed. I do not see a separate flaw in the architecture or loss that would make the method impossible; the concern is about evidence, not mechanism. The paper's qualitative examples and downstream detection table are useful but do not cover degraded test cases (detection is only on MSRS clean inputs), so Table III carries the full weight. Thus the correct disposition remains CONDITIONAL: the method could be fine, but the central quantitative claim should be re-verified with fidelity-aware metrics. No verdict change from the reader's conditional is needed.","tokens_in":22110,"tokens_out":4495,"duration_ms":50299,"concrete_test":"On the same EMS test splits used in Table III, apply GD2Fusion and the seven comparison pipelines exactly as in Sec. IV-D. Compute PSNR and SSIM (and, if feasible, LPIPS) between each fused output and the degradation-free references I_ref_ir and I_ref_vi used in Eqs. 10-12, plus the max-intensity target max(I_ref_ir, I_ref_vi) from Eq. 10. Report per-degradation-combination means and standard deviations. If GD2Fusion does not rank first on these fidelity metrics, Table III's evidence for dual-source degraded superiority is an artifact of gradient/contrast metrics rather than restoration quality.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—'consistently outperforms all competing methods' in dual-source degraded scenarios (Sec. IV-D, Table III)—is supported only by AG, EI, SD, and SF (Sec. IV-B). These are gradient/contrast/edge-energy statistics that reward over-enhancement: a fusion that amplifies high-frequency content and boosts contrast can raise all four while degrading fidelity or introducing artifacts. The paper's own training objective (Eqs. 10-12) supervises against 'degradation-free infrared and visible reference images' I^ref_ir and I^ref_vi from the EMS dataset, so such references exist for the training set; Sec. IV-B nonetheless rules out reference-based metrics for the degraded evaluation. For the EMS test splits, if the same clean references are available, Table III can be checked directly. If they are not available, then the decision to omit any fidelity-oriented metric leaves the headline superiority claim untested except by metrics that the method's own loss is designed to maximize. Because this table is the only quantitative support for the central degraded-scenario claim (Fig. 7 is selective, and no code or exact prompts are provided), the claim is not currently secured.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GD2Fusion, an end-to-end infrared-visible image fusion framework for dual-source degraded scenarios. It uses frozen CLIP text features (prompts) to inject degradation-awareness into two complementary branches: a frequency-domain GFMSE module for modality-specific degradation suppression and feature extraction, and a spatial-domain GSMAF module for cross-modal degradation filtering and adaptive fusion. The network is trained with intensity, texture, and color losses against degradation-free reference images from the EMS dataset, and evaluated on EMS degraded test scenarios, MSRS/LLVIP high-quality scenarios, and a downstream object-detection task. The central claim is that GD2Fusion outperforms existing fusion algorithms and pre-enhancement-plus-fusion strategies in dual-source degraded scenarios.","tokens_in":22340,"tokens_out":5329,"duration_ms":57633,"significance":"If the central claim is secured, the paper would make a useful contribution: it couples VLM-based degradation perception with frequency/spatial dual-domain modeling in a single end-to-end framework, avoiding the error accumulation of cascaded pre-enhancement. The architecture is clearly described, the high-quality-input experiments (Table II) show competitive or superior performance against seven strong baselines with reference-based metrics, and the downstream detection experiment (Table V) is a positive addition. However, the paper's headline contribution—robust fusion under dual-source degradation—rests on a small set of no-reference sharpness/contrast metrics, and the manuscript contains an internal contradiction about the availability of reference images. The exact VLM prompts are not disclosed, and the most related multi-degradation baselines (DRMF, TFD2Fusion) are omitted from the experiments. These issues are load-bearing for the claimed contribution and need to be addressed.","major_comments":[{"comment":"Section IV-B states that reference-based metrics cannot be applied because no ideal ground-truth images exist. Yet Eqs. (10)-(12) define the training loss using I_ref_ir and I_ref_vi as 'degradation-free infrared and visible reference images' obtained, per Sec. IV-A, from the EMS dataset. Since the same EMS dataset supplies the degraded test pairs, these clean references are available for the test splits. The claim in Sec. IV-B is therefore internally inconsistent. This matters because Table III is the sole quantitative support for the abstract's claim of superiority in dual-source degraded scenarios, and its four metrics are gradient/contrast statistics. Please report reference-based metrics (e.g., FMI, MI, VIF, Qabf, Qw, Qy, as in Table II) on the EMS test splits, or explain precisely why the training references cannot be used for evaluation.","section":"Sec. IV-B vs. Eqs. (10)-(12)"},{"comment":"All degraded-scenario quantitative conclusions rest on AG, EI, SD, and SF. These no-reference metrics measure average gradient, edge intensity, standard deviation, and spatial frequency; they reward increased high-frequency content and contrast and can be inflated by over-enhancement or artifacts. The proposed loss (Eqs. 10-12) explicitly supervises intensity, gradient, and color against per-pixel maxima of the source references, so the network is trained to maximize essentially the same quantities later used to declare superiority. At minimum, add fidelity-oriented no-reference or perceptual metrics (e.g., NIQE, PIQE, BRISQUE) and/or a human study for the degraded scenario; otherwise the statement in Sec. IV-D that GD2Fusion 'consistently outperforms all competing methods across all assessed metrics' is not sufficient to support the paper's central claims.","section":"Table III and Table IV"},{"comment":"The exact textual prompts P_ir and P_vi—the core VLM guidance—are never listed. Fig. 2 shows only a schematic \"Prompt\"; the strings are not recoverable from the text, and code is promised only after acceptance. Since the prompts are hand-crafted for the four degradation types in EMS (Sec. IV-A), their exact content is essential for reproducibility and for assessing the generality of the approach. Please include the prompt templates in the paper or supplementary material and release code at review time. In addition, the most closely related recent methods for multi-degradation fusion, DRMF [28] and TFD2Fusion [40], are discussed in Sec. II but do not appear in Tables I-III; include them in the comparison or provide a clear justification for their exclusion.","section":"Sec. III-A / Sec. IV-A"}],"minor_comments":[{"comment":"The second assignment to F_local should presumably be F_global = T rmBlock^M(F_guided); currently the equation overwrites F_local, which conflicts with the use of both F_local and F_global in Eq. (4).","section":"Eq. (9)"},{"comment":"\"standard division\" should be \"standard deviation\"; the citation [53] (Bragg grating sensors) appears unrelated to this metric and should be replaced.","section":"Sec. IV-B"},{"comment":"There are typographical inconsistencies: Index Terms contain \"Languge\"; Sec. III-C title uses \"Model-Aggregated\" while the text uses \"Modality-Aggregated\"; the method name is rendered inconsistently as \"GD2Fusion\" and \"GD 2Fusion\". Please unify and proofread.","section":"Throughout"},{"comment":"State explicitly that the 69/69/35/35 EMS test pairs are disjoint from the 2,278 training pairs, and describe how the degradation-free reference images used in the loss were constructed. This is needed to rule out train/test leakage and to clarify the contradiction flagged in the major comments.","section":"Sec. IV-A"},{"comment":"The displayed prompt text is illegible. Provide the exact infrared and visible prompt strings in the caption or in a supplementary table so readers can verify the degradation-awareness mechanism.","section":"Fig. 7"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is valid and lands directly: the training loss uses degradation-free references from EMS while Sec. IV-B claims such references do not exist, and the degraded-scenario evaluation uses only four no-reference sharpness/contrast metrics. The paper's core contribution is currently supported mainly by Table III, so this must be fixed before the claim of superiority is accepted. The omission of DRMF and TFD2Fusion from the benchmarks is also conspicuous given that they are discussed in the related work. If the evaluation is strengthened, the paper could be a solid contribution to the fusion literature; the architecture and high-quality-input results are promising."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a well-put-together architecture paper for a real problem—IVIF when both modalities are degraded—but the headline quantitative claim for the degraded scenario doesn't hold up yet. The evaluation logic has an internal inconsistency that needs fixing before I'd trust the superiority claim.\n\nThe genuinely new thing is the combination: frozen CLIP text prompts for degradation perception, affine-conditioning of wavelet sub-bands (GFMSE), plus a prompt-aggregated spatial fusion branch (GSMAF), all integrated end-to-end. That combination isn't in Text-IF or TFD2Fusion, and the paper does a decent job of positioning itself against them. The ablations are thorough—module removals, depth, loss terms—and the high-quality-input results on MSRS/LLVIP (Table II) are strong across reference-based metrics. The downstream detection evaluation is a nice addition.\n\nThe soft spots are mostly about the degraded-scenario evaluation, which is the paper's raison d'être. Section IV-B says reference-based metrics can't be used because no degradation-free ground truth exists, yet the training loss (Eqs. 10-12) supervises against degradation-free references I_ref_ir and I_ref_vi from the same EMS dataset. You can't have it both ways: if those references exist for training, they almost certainly exist for the held-out test splits, and Table III should be checked against SSIM/PSNR or at least a fidelity metric. As it stands, the claim of 'consistently outperforms' rests entirely on AG/EI/SD/SF, which are gradient and contrast statistics that an over-sharpened or over-contrasted output will inflate. That's not a fatal flaw—the fixes are straightforward—but it's a load-bearing gap.\n\nTwo more things: the paper doesn't say whether the 2,278 training pairs and the 208 test pairs from EMS are disjoint, and there are no error bars or significance tests. And no code or exact prompt text is provided, which compounds the reproducibility issue since the prompts are hand-designed for exactly the four degradations in the test set. Those are addressable.\n\nOverall: the architecture is plausible and the high-quality-input evidence is substantive. The degraded-scenario claim is not yet secured. This deserves peer review—an editor should send it out—but the revisions needed are more than cosmetic. Who it's for: people working on fusion under degradation, or on VLM-guided low-level vision; they'll get a useful architectural baseline and a good case study in why metric choice matters. I'd bring it to a reading group to argue about the metrics, but I wouldn't cite it in its current form.","headline":"A plausible architecture with a real niche, but the degraded-scenario superiority claim is not supported by the reported metrics and needs an internal-contradiction fix before I'd trust it.","tokens_in":22904,"tokens_out":5238,"would_cite":false,"duration_ms":51854,"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":"One network fuses degraded IR and visible images end-to-end, no pre-enhancement stage needed.","keywords":["infrared-visible image fusion","degradation-aware fusion","vision-language model","frequency domain","wavelet transform","dual-domain fusion","low-light image enhancement","object detection"],"falsifier":"A reference-based evaluation on held-out degradation-free pairs from the same EMS data—comparing GD2Fusion against cascaded IAT/SwinIR/MIRNet-plus-fusion pipelines—that fails to rank GD2Fusion first, or a perceptual study in which human raters prefer the cascaded outputs, would settle the claim.","tokens_in":21938,"feed_emoji":"🖼️","tokens_out":6054,"duration_ms":59040,"temperature":0.7,"pith_summary":"This paper tries to establish that a single end-to-end network can fuse infrared and visible images when both sources are degraded—low-light or overexposed visible images paired with low-contrast or noisy infrared images—without first running separate enhancement steps. The proposed GD2Fusion system uses text prompts from a frozen vision-language model to tell the network what degradations are present, then processes the two modalities in both frequency and spatial domains. The authors report that on the EMS benchmark, GD2Fusion outperforms seven prior methods, including cascaded pre-enhancement-plus-fusion strategies, on all four no-reference metrics across four degradation combinations. If correct, the framework would remove the error-accumulating pre-enhancement stage that current degraded fusion pipelines depend on, and make fusion more practical for real-world low-light and noisy imaging.","feed_headline":"One network fuses degraded IR and visible images end-to-end","feed_subtitle":"Text prompts guide frequency and spatial branches to suppress low light, noise, overexposure, and low contrast together.","key_machinery":"The load-bearing mechanism is prompt-guided affine modulation in two complementary domains. GFMSE applies a Haar discrete wavelet transform to each modality, separates the low-frequency approximation (luminance and contrast) from high-frequency details (texture and noise), and uses affine parameters generated from modality-specific prompt features to re-weight those sub-bands before convolutional and Transformer blocks. GSMAF concatenates the two modalities' prompt features into an aggregated guidance vector, applies channel modulation and affine transformation to the fused spatial features, then aggregates them through multi-scale convolutions with kernel sizes 3, 5, and 7 plus Transformer","core_discovery":"The central claim is that degradation perception and fusion should be optimized in one framework rather than staged. GD2Fusion feeds textual prompts through a frozen vision-language text encoder, producing per-modality guidance features that are injected as affine modulation into two parallel pathways: the Guided Frequency Modality-Specific Extraction (GFMSE) module, which decomposes each modality with a discrete wavelet transform and suppresses degradation in low- and high-frequency sub-bands separately, and the Guided Spatial Modality-Aggregated Fusion (GSMAF) module, which filters cross-modal degradation and aggregates multi-source spatial features. The frequency features are injected int","pith_inferences":["If the mechanism is as general as it appears, the same prompt-guided dual-domain design could be extended to other paired-modality degradations—for example haze or rain in the visible image with thermal blur in the infrared—simply by writing new prompts and training on paired degraded examples.","The training loss uses degradation-free reference images, while the degraded-scenario evaluation relies on no-reference metrics; a held-out reference-based evaluation against those clean counterparts would directly show whether the quantitative gains reflect true fidelity rather than contrast amplification.","Since the text encoder is frozen and prompts are plain text, the framework could in principle be driven by an automatic captioner that detects degradations and writes the prompt at runtime, removing the manual selection the paper criticizes."],"forward_implications":["Fusion systems can drop the manual pre-enhancement stage for common dual degradations; one trained model handles low-light or overexposed visible input together with low-contrast or noisy infrared input.","Changing the text prompt changes the degradation guidance, so the same network can be steered toward different degradation types without retraining the fusion backbone.","Because wavelet sub-bands isolate luminance and contrast problems from noise and texture problems, the architecture can suppress each degradation class in the frequency band where it lives rather than treating the whole image uniformly.","Downstream tasks benefit directly: on MSRS, GD2Fusion's fused outputs give YOLOv5 higher mAP than any of the seven comparison methods, suggesting better inputs for detection.","The decoupled cascade of enhancement-then-fusion, with its own objective and feature space, is not necessary for state-of-the-art degraded fusion."],"supporting_citations":[{"why":"Supplies the EMS degraded-pair dataset, the Text-IF baseline, and the degradation-free reference images used in training.","marker":"[20]"},{"why":"Provides the frozen vision-language text encoder that turns degradation prompts into guidance features.","marker":"[36]"},{"why":"Supplies the discrete wavelet transform and inverse transform that separate low- and high-frequency sub-bands in GFMSE.","marker":"[41]"},{"why":"IAT: the visible-light brightness-adjustment pre-enhancement baseline used in the fair degraded-scenario comparison.","marker":"[17]"},{"why":"SwinIR: the infrared denoising baseline used in the degraded-scenario comparison.","marker":"[18]"},{"why":"MIRNet: the infrared contrast-enhancement baseline used in the degraded-scenario comparison.","marker":"[61]"},{"why":"Supplies the high-quality training set for the input-assumption model and a second evaluation dataset.","marker":"[45]"}],"fun_headline_variants":["Text prompts guide dual-domain fusion for degraded IR and visible images","VLM-guided fusion suppresses noise, low light, and overexposure together","One-pass fusion handles degraded infrared and visible inputs","Degradation-aware fusion: text-guided frequency and spatial branches","Fuse degraded IR and visible images in one pass with text guidance"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The degraded-scenario superiority rests on the assumption that average gradient, edge intensity, standard deviation, and spatial frequency faithfully measure fusion quality—four no-reference metrics that can reward amplified contrast or noise as much as genuine restoration.","fun_headline_variants_meta":{"raw":{"variants":["Text prompts guide dual-domain fusion for degraded IR and visible images","VLM-guided fusion suppresses noise, low light, and overexposure together","One-pass fusion handles degraded infrared and visible inputs","Degradation-aware fusion: text-guided frequency and spatial branches","Fuse degraded IR and visible images in one pass with text guidance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001029,"raw_usage":{"total_tokens":4164,"prompt_tokens":727,"completion_tokens":3437,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":3351}},"tokens_in":471,"tokens_out":3437,"duration_ms":22053,"temperature":1.0,"reasoning_tokens":3351,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:41:35.792654+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reference-based evaluation on held-out degradation-free pairs from the same EMS data—comparing GD2Fusion against cascaded IAT/SwinIR/MIRNet-plus-fusion pipelines—that fails to rank GD2Fusion first, or a perceptual study in which human raters prefer the cascaded outputs, would settle the claim.","supporting_citations":[{"cited_title":"Text- if: Leveraging semantic text guidance for degradation- aware and interactive image fusion,","cited_arxiv_id":null,"evidence_quote":"Supplies the EMS degraded-pair dataset, the Text-IF baseline, and the degradation-free reference images used in training."},{"cited_title":"Learning transferable visual models from natural language supervision,","cited_arxiv_id":null,"evidence_quote":"Provides the frozen vision-language text encoder that turns degradation prompts into guidance features."},{"cited_title":"The haar wavelet transform: its status and achievements,","cited_arxiv_id":null,"evidence_quote":"Supplies the discrete wavelet transform and inverse transform that separate low- and high-frequency sub-bands in GFMSE."},{"cited_title":"You only need 90k parameters to adapt light: a light weight transformer for image enhancement and exposure correction,","cited_arxiv_id":null,"evidence_quote":"IAT: the visible-light brightness-adjustment pre-enhancement baseline used in the fair degraded-scenario comparison."},{"cited_title":"Swinir: Image restoration using swin trans- former,","cited_arxiv_id":null,"evidence_quote":"SwinIR: the infrared denoising baseline used in the degraded-scenario comparison."},{"cited_title":"Learning enriched features for fast image restoration and enhancement,","cited_arxiv_id":null,"evidence_quote":"MIRNet: the infrared contrast-enhancement baseline used in the degraded-scenario comparison."},{"cited_title":"Llvip: A visible-infrared paired dataset for low-light vision,","cited_arxiv_id":null,"evidence_quote":"Supplies the high-quality training set for the input-assumption model and a second evaluation dataset."}],"review_version":1}