{"id":"dea8f446-2482-487e-9d52-11c6c0be7623","arxiv_id":"2504.19127","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"DeepSPG combines Retinex decomposition with image-level semantic features from HRNet and text-level CLIP alignment to improve low-light image enhancement, reporting state-of-the-art PSNR/SSIM on LOL-v1, LOL-v2-synthetic, and SID.","lead":"This paper describes a low-light image enhancement system that uses both an image segmentation model and a language model to guide brightening, instead of only learning a direct pixel-to-pixel mapping. It reports higher quality scores than several existing methods on five benchmarks, while using far fewer parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's FLOPs/params for DeepSPG omit the frozen HRNet-W48 segmentation backbone that runs at inference; with it counted, DeepSPG-large is likely larger and heavier than Retinexformer, so the claimed efficiency trade-off is unsupported.","rationale":"The reader's weakest assumption was about the semantic accuracy of frozen HRNet features on under-exposed, noisy inputs. I agree that this is a real limitation, and the paper's Sec. 5 admits it. However, the more decisive problem for the paper's stated central claim is that the efficiency comparison in Table 1 appears to count only the trainable refinement network. Because HRNet-W48 is executed at inference to obtain the image-level semantic prior, omitting its cost makes the 'better trade-off' claim not comparable to the listed baselines. The paper itself is internally inconsistent in another way: the SOTA numbers come from DeepSPG-large (2.13M params, 26.03 GFLOPs), while the '0.55M params / 6.91 GFLOPs' efficiency advantage belongs to the base model, which is not SOTA on LOL-v2-real or SMID in Table 1 (e.g., 22.78 vs 22.80 on LOL-v2-real, 28.15 vs 29.15 on SMID). This does not invalidate the enhancement architecture, but it means the central contribution as written is overclaimed. The concern is concrete and checkable: profiling the full inference graph including HRNet-W48 will settle it. This supports keeping the reader's CONDITIONAL verdict: the paper should be accepted only after the complexity reporting and the SOTA/efficiency claims are corrected and verified.","tokens_in":13267,"tokens_out":7110,"duration_ms":73826,"concrete_test":"Recompute Table 1's complexity columns for DeepSPG and DeepSPG-large by profiling the full inference graph on a LOL-v1 image at the resolution used in Table 1, including the frozen HRNet-W48 forward pass needed to generate M_seg and X_seg, and also measuring the CLIP text/visual encoders if they are retained at inference. Compare the resulting FLOPs and parameter counts with Retinexformer measured under identical settings. If the end-to-end counts exceed about 1.61M params or 15.57 GFLOPs, the efficiency trade-off claim fails as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Sec. 1 is that DeepSPG 'achieves SOTA performance on five widely used datasets and a better trade-off between performance and computational costs.' The efficiency half of this claim rests on Table 1, which lists DeepSPG at 6.91 GFLOPs/0.55M params and DeepSPG-large at 26.03 GFLOPs/2.13M params. However, the image-level semantic prior (Sec. 3.3) requires running a frozen HRNet-W48 segmentation model pre-trained on ADE20K on every input to produce M_seg and X_seg (Eq. 1), and the semantic loss (Eq. 13) requires a forward pass through HRNet during training. HRNet-W48 has on the order of 60-70M parameters and, depending on input resolution, tens to hundreds of GFLOPs. The table appears to report only the trainable refinement network. Counting the frozen backbone, DeepSPG-large's deployment cost exceeds Retinexformer (1.61M params, 15.57 GFLOPs), and the '0.55M parameters' statement is not a fair comparison to baselines that do not use a pretrained segmentation network. Thus the 'better trade-off' contribution is not established. The paper should either count the full inference graph or restrict the claim to trainable parameters and benchmark end-to-end latency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes DeepSPG, a low-light image enhancement framework built on Retinex decomposition that incorporates image-level semantic priors from a frozen HRNet-W48 segmentation model and text-level priors from CLIP. The framework uses a semantic-aware embedding module (SEM) to fuse hierarchical semantic features into the reflectance branch and a multimodal loss to align enhanced images with text prompts. Experiments are reported on five benchmark datasets (LOL-v1, LOL-v2-real, LOL-v2-syn, SID, SMID) with quantitative comparisons against eleven methods, ablations, and visual results. The authors claim state-of-the-art performance and better computational efficiency.","tokens_in":13557,"tokens_out":8488,"duration_ms":71428,"significance":"If the results are taken at face value, DeepSPG provides a useful demonstration that high-level semantic priors can improve low-light enhancement, especially in dark, information-depleted regions. The paper ships public code, which supports reproducibility, and the ablation study provides evidence that both image-level and text-level priors individually contribute. However, the strength of the empirical claim is weakened by several discrepancies between the reported numbers and the statements made in the text, and by the omission of the frozen backbone from the efficiency accounting. The core idea is interesting and the experimental structure is sound; the issues are mostly in the reporting and verification, though they are load-bearing for the SOTA and efficiency claims.","major_comments":[{"comment":"The caption says 'Our DeepSPG significantly outperforms SOTA algorithms,' but the table does not support this for the base model: DeepSPG trails Retinexformer on LOL-v2-real (22.78 vs 22.80 PSNR) and SMID (28.15 vs 29.15 PSNR), and ties on SID (24.44 vs 24.44). DeepSPG-large also trails Retinexformer on SMID (29.08 vs 29.15 PSNR). The sentence 'achieving PSNR improvement in three datasets (LOL-v1, LOL-v2-syn, and SID)' is therefore incorrect for SID. Please correct these claims and either reposition the contribution as competitive rather than SOTA, or conduct additional experiments to resolve the discrepancies.","section":"Sec. 4.2, Table 1"},{"comment":"The complexity numbers in Table 1 (6.91 GFLOPs, 0.55M params for DeepSPG; 26.03 GFLOPs, 2.13M for DeepSPG-large) count only the trainable subnetwork. At inference the image-level prior requires a forward pass of the frozen HRNet-W48 segmentation model (Eq. 1), which has tens of millions of parameters and substantial FLOPs. Without counting this backbone, the claim of a 'better trade-off between performance and computational costs' is unsupported and is not an apples-to-apples comparison with baselines that do not use such a backbone. Please report end-to-end inference cost (params, FLOPs, and latency) for the full model, or clearly and consistently state that the reported numbers are for the trainable part only, and adjust the efficiency claim accordingly.","section":"Sec. 3.3, Eq. (1); Sec. 4.2, Table 1"},{"comment":"The coarse-to-fine (C2F) training scheme, which the ablation credits with a 0.28 dB PSNR improvement, is never defined anywhere in the paper. It appears only in Table 2 and in the caption of Fig. 2. Without a description of the scheme (e.g., whether it refers to progressive resolution training, curriculum over stages, or something else), the ablation is not reproducible and the contribution of this component cannot be assessed. Please add a precise definition and implementation details.","section":"Sec. 4.3, Table 2"},{"comment":"The semantic loss is not clearly specified. Eq. (1) defines X_seg as the segmentation map of the low-light input, while the text in Sec. 3.2 says the loss is computed 'between the generated and ground-truth segmentation maps.' Eq. (13) writes Lsem = KL(X_seg || \\hat X_seg) without defining \\hat X_seg. If X_seg is the input's segmentation, the asymmetric KL would pull the enhanced image's semantic distribution toward the low-light input's distribution, which may be wrong in exactly the noise-disturbed cases the paper acknowledges in Sec. 5. Please define \\hat X_seg explicitly (e.g., segmentation of the enhanced image or of the ground truth), state the direction of the KL divergence used in practice, and justify why this choice is beneficial rather than harmful.","section":"Sec. 3.5, Eq. (13)"},{"comment":"SKF [14] is the most closely related semantic-guided LLIE method and provides the SEM module that DeepSPG adopts. Yet SKF is not included in the quantitative comparison in Table 1. Since the paper claims to surpass previous methods and specifically to improve on 'image-level semantics without multimodal learning,' the omission of SKF leaves the advantage over the direct predecessor undemonstrated. Please add SKF to the benchmarks or explain why it cannot be compared.","section":"Sec. 2.2, Sec. 4.2"},{"comment":"The architecture of DeepSPG-large is never described. Table 1 reports results for DeepSPG-large and the text uses them to support the SOTA claim, but the paper does not state how DeepSPG-large differs from the base model (e.g., width, depth, or input resolution). Without this information, the large-variant results are not reproducible. Please add a description of the variant or remove it from the main comparison.","section":"Sec. 4.1, Table 1"}],"minor_comments":[{"comment":"The caption of Fig. 4 says 'our PiCat' where it should say 'our DeepSPG.' Please correct this typo, which appears to be a leftover from another manuscript.","section":"Fig. 4 caption"},{"comment":"The efficiency comparison mixes metrics: '2.3 times lighter' refers to FLOPs (6.91 vs 15.57 GFLOPs), while '113 times more efficient' refers to FLOPs of MIRNet, but the parameter counts give different ratios (0.55/1.61 is about 2.9x). Please state which quantity is meant and report both params and FLOPs consistently.","section":"Sec. 4.2"},{"comment":"The values of the loss weights λ1, λ2, λ3, and λ4 are not reported, even though they are said to be 'empirically tuned.' Please include the exact values used in the experiments.","section":"Sec. 3.5"},{"comment":"The name 'Yann LeCun' is misspelled as 'Yan LeCun' in the quotation at the start of the introduction.","section":"Sec. 1"},{"comment":"The paper contains leftover ACM template artifacts, including 'Conference acronym 'XX,' the ACM copyright block, and '2018' in the reference format. These should be removed or updated for a clean submission.","section":"Throughout"},{"comment":"The terms 'reflection' and 'reflectance' are used interchangeably (e.g., 'reflection maps' vs 'reflectance map'). Please standardize the terminology.","section":"Sec. 3.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be an early preprint (with ACM template artifacts and a caption typo referencing 'PiCat'). The self-citations [21] and [41] are from the same group; while they are not load-bearing, the editor may wish to check for any undisclosed joint work. The central concern is that the reported numbers do not support the stated SOTA claim on two datasets and the efficiency claim is not additive for the full inference graph; these are fixable in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent, honest engineering paper. The specific combination—Retinex decomposition, frozen HRNet-W48 semantic features, and CLIP text alignment—is not in the cited literature, and the ablations give real evidence that each piece helps: +1.06 dB PSNR from the image-level prior, +0.34 dB from text, +0.28 dB from the C2F scheme. Code is promised, the failure case is shown, and the limitation section admits the core weakness. Credit where due.\n\nSoft spots, in order of severity. First, the efficiency claim is misleading. Table 1 lists DeepSPG at 0.55M params and 6.91 GFLOPs, but the frozen HRNet-W48 segmentation model runs on every input to produce the semantic features and segmentation map. HRNet-W48 is roughly 60–70M params and tens to hundreds of GFLOPs depending on input size. Count the full inference graph and DeepSPG-large is likely heavier and slower than Retinexformer, so the 'better trade-off' claim in Sec. 1 is unsupported. The stress-test note is correct.\n\nSecond, the SOTA claim overreaches. In Table 1, base DeepSPG trails Retinexformer on LOL-v2-real (22.78 vs 22.80) and SMID (28.15 vs 29.15), and even DeepSPG-large trails on SMID (29.08 vs 29.15). The abstract's 'superior performance across five benchmark datasets' is not what their own table shows. They should descope to the datasets where they actually win.\n\nThird, the C2F training scheme is ablated in Tab. 2 but never defined in the method. Loss weights lambda1–4 are unreported, error bars are absent, and Fig. 4's caption says 'our PiCat' instead of DeepSPG—an obvious copy-paste artifact. These are fixable, but they make the paper feel rushed.\n\nOne more point: the semantic loss and the semantic prior both use the same HRNet, so the gain could partly reflect a stronger feature extractor rather than 'semantic guidance' per se. Not fatal, but worth an ablation or a sentence of discussion.\n\nWho is this for? Someone working on low-light enhancement who wants a competent baseline with multimodal priors and a fairly clean ablation. It deserves a serious referee, not a desk reject. I'd send it to review, expecting heavy revision of the efficiency accounting, the SOTA claim, and the missing training details. I wouldn't cite it in my own work in the next year—not my area, and the numbers need fixing first. Reading group: maybe.","headline":"A solid LLIE engineering contribution whose real claims—semantic guidance helps, compact trainable model—are undercut by an overstated SOTA claim and an efficiency comparison that omits the frozen HRNet backbone.","tokens_in":14135,"tokens_out":2265,"would_cite":false,"duration_ms":20967,"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":"DeepSPG claims that steering Retinex reflectance refinement with image-level segmentation features and text-level CLIP alignment improves low-light enhancement, reporting state-of-the-art PSNR/SSIM on four of five benchmarks with 0.55M…","keywords":["low-light image enhancement","Retinex decomposition","semantic prior guidance","multimodal learning","CLIP alignment","semantic segmentation features","reflectance refinement","image restoration"],"falsifier":"Take the darkest, noisiest quartile of SID test images, run the frozen segmentation model on the raw low-light inputs, and compare its output against ground-truth segmentation. If PSNR stays high even where the segmentation map is wrong, while removing the semantic module barely changes performance, then the image-level prior is not load-bearing; the paper's own reported failure case is the natural starting point.","tokens_in":13006,"feed_emoji":"💡","tokens_out":17775,"duration_ms":152886,"temperature":0.7,"pith_summary":"DeepSPG sets out to show that a low-light enhancement network recovers more detail when it is told what the scene contains, not just how dark it is. The paper builds on Retinex decomposition, separating each image into illumination (brightness) and reflectance (object colors and textures independent of lighting), $X_{in}=X_l\\cdot X_r$, and refines the reflectance branch under two semantic priors: hierarchical features from a frozen semantic segmentation network (HRNet-W48), and text-level constraints from CLIP, a pretrained vision-language model, that pull the enhanced output toward the description \"high-light image\" and away from \"low-light image.\" In the reported experiments, the large variant records the highest PSNR on four of the five tested datasets (LOL-v1, LOL-v2-real, LOL-v2-synthetic, SID), while the base model runs at 0.55M parameters, and the ablations attribute gains of about 1.7 dB and 0.022 SSIM to adding the two semantic priors plus coarse-to-fine training. If the claim holds, it means high-level semantic knowledge is not decoration but a practical way to stabilize color and structure in the darkest parts of an image.","feed_headline":"Scene and text cues push low-light enhancement past previous best","feed_subtitle":"The method pairs scene understanding with text constraints to refine image reflectance, beating earlier methods on four benchmarks.","key_machinery":"The machinery is Retinex decomposition plus two aligned semantic streams. The governing identity is $X_{in}=X_l\\cdot X_r$: the illumination map carries brightness and is adjusted, while the reflectance map carries object color and structure and is refined under semantic guidance. The fusion point is the semantic-aware embedding module (SEM), which projects reflectance and segmentation features into key/query/value spaces, computes the softmax correlation $A^b=\\mathrm{Softmax}(W_k(F_i^b)\\times W_q(F_s^b)/\\sqrt{C})$, and adds a residual feed-forward refinement. The text stream adds a cosine-similarity loss $L_{mul}=D_{cos}(X_{out},T_L)-D_{cos}(X_{out},T_H)$ that separates the enhanced image from the \"low-light\" prompt and toward the \"high-light\" prompt, and the full objective is a weighted sum of pixel, edge, semantic-KL, and multimodal losses over a coarse-to-fine multi-scale structure.","core_discovery":"The central claim is that semantic priors can be made load-bearing in low-light enhancement by attaching them to the reflectance map rather than to the final image alone. After Retinex decomposition, DeepSPG feeds the reflectance branch with multi-scale features and a segmentation map from a frozen HRNet-W48 segmentation model, using a semantic-aware embedding module whose attention correlation reweights reflectance features by their similarity to semantic prototypes. A CLIP-based text loss then pushes the enhanced image's embedding toward the \"high-light image\" prompt and away from \"low-light image,\" while a KL-divergence semantic loss keeps the enhanced image's segmentation map aligned with the ground truth's. On the paper's numbers, this combination gives DeepSPG-large 27.03 dB PSNR on LOL-v1 and 28.08 dB on LOL-v2-synthetic, surpassing the previous best by 1.87 dB and 2.41 dB respectively, and the base model reports 26.87 dB and 27.11 dB on the same two datasets with only 0.55M parameters.","pith_inferences":["A testable extension the paper leaves implicit is that swapping in a noise-robust segmentation model should transfer directly to gains in extreme dark scenes, since the paper's own failure case is attributed to wrong segmentation features under noise.","The text prior is a single global contrast between \"low-light image\" and \"high-light image\"; per-scene prompts describing the dominant objects could give the vision-language loss finer control than the current pair.","The semantic loss compares segmentation maps, so the benefit may be concentrated in scenes with recognizable objects such as people, text, and indoor objects; on textureless or uniformly dark regions, pixel and edge losses may carry most of the weight, a split the paper does not measure."],"forward_implications":["In the reported ablations, adding the image-level semantic prior raises PSNR by 1.06 dB and SSIM by 0.013, and adding the text-level prior adds another 0.34 dB and 0.003, so each modality contributes measurably to the final result.","The large variant's reported 27.03 dB on LOL-v1 and 28.08 dB on LOL-v2-synthetic exceed the strongest prior baseline by 1.87 dB and 2.41 dB, indicating that the semantic-prior design closes a real gap on paired benchmarks.","With 0.55M parameters and 6.91 G FLOPs, the base model is lighter than the transformer baselines it is compared with, so semantic guidance is compatible with efficient deployment.","Because both semantic priors come from frozen pretrained models, the framework inherits their limitations: the paper states that when the segmentation model misreads noisy dark input, enhancement fails to recover correct semantics and colors."],"supporting_citations":[{"why":"Supplies the Retinex image-formation model that the reflectance/illumination decomposition rests on.","marker":"[15]"},{"why":"Introduces the semantic-aware embedding module that this paper adapts to fuse segmentation features into the reflectance branch.","marker":"[14]"},{"why":"Provides the frozen high-resolution segmentation network whose features and maps carry the image-level prior.","marker":"[29]"},{"why":"Provides the joint image-text embedding space and text encoder used by the multimodal loss.","marker":"[28]"},{"why":"Defines the LOL-v1 paired low-light benchmark used for training and evaluation.","marker":"[31]"},{"why":"Defines the LOL-v2 real and synthetic benchmarks used to measure generalization.","marker":"[32]"},{"why":"Defines the SID extreme low-light dataset used for real-world dark-scene evaluation.","marker":"[33]"},{"why":"Defines the SMID dark-scene dataset used for one of the five benchmark evaluations.","marker":"[34]"},{"why":"Is the strongest PSNR baseline (Retinexformer) that DeepSPG compares against and claims to surpass on several datasets.","marker":"[4]"}],"fun_headline_variants":["Semantic priors on reflectance map beat prior low-light methods","Image and text semantics fuse to enhance low-light photos","DeepSPG uses multimodal cues to outperform previous low-light models","Semantic segmentation and CLIP guide low-light enhancement","Semantic-aware low-light enhancement surpasses prior art"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a segmentation network trained on ordinary well-lit images can still read what objects are in an extremely dark, noisy photo; if it cannot, both the image-level guidance and the semantic loss are steering the enhancement with wrong information.","fun_headline_variants_meta":{"raw":{"variants":["Semantic priors on reflectance map beat prior low-light methods","Image and text semantics fuse to enhance low-light photos","DeepSPG uses multimodal cues to outperform previous low-light models","Semantic segmentation and CLIP guide low-light enhancement","Semantic-aware low-light enhancement surpasses prior art"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1560,"prompt_tokens":989,"completion_tokens":571,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":491}},"tokens_in":605,"tokens_out":571,"duration_ms":5718,"temperature":1.0,"reasoning_tokens":491,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T06:00:18.973499+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the darkest, noisiest quartile of SID test images, run the frozen segmentation model on the raw low-light inputs, and compare its output against ground-truth segmentation. If PSNR stays high even where the segmentation map is wrong, while removing the semantic module barely changes performance, then the image-level prior is not load-bearing; the paper's own reported failure case is the natural starting point.","supporting_citations":[{"cited_title":"The retinex theory of color vision","cited_arxiv_id":null,"evidence_quote":"Supplies the Retinex image-formation model that the reflectance/illumination decomposition rests on."},{"cited_title":"Learning semantic-aware knowledge guidance for low-light image enhancement","cited_arxiv_id":null,"evidence_quote":"Introduces the semantic-aware embedding module that this paper adapts to fuse segmentation features into the reflectance branch."},{"cited_title":"Deep high- resolution representation learning for visual recognition","cited_arxiv_id":null,"evidence_quote":"Provides the frozen high-resolution segmentation network whose features and maps carry the image-level prior."},{"cited_title":"Deep retinex decom- position for low-light enhancement","cited_arxiv_id":null,"evidence_quote":"Defines the LOL-v1 paired low-light benchmark used for training and evaluation."},{"cited_title":"Sparse gradient regularized deep retinex network for robust low-light image enhancement","cited_arxiv_id":null,"evidence_quote":"Defines the LOL-v2 real and synthetic benchmarks used to measure generalization."},{"cited_title":"Learning to see in the dark","cited_arxiv_id":null,"evidence_quote":"Defines the SID extreme low-light dataset used for real-world dark-scene evaluation."},{"cited_title":"Seeing motion in the dark","cited_arxiv_id":null,"evidence_quote":"Defines the SMID dark-scene dataset used for one of the five benchmark evaluations."},{"cited_title":"Retinexformer: One-stage retinex-based transformer for low-light image en- hancement","cited_arxiv_id":null,"evidence_quote":"Is the strongest PSNR baseline (Retinexformer) that DeepSPG compares against and claims to surpass on several datasets."}],"review_version":1}