{"id":"0fc96bf2-4663-4a13-ad71-5543b4b7c664","arxiv_id":"2412.11883","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"AllSky, a U-Net trained on HDR sky photos with cascade exposure losses and a learned LDR-to-EDR head, improves sun-region dynamic range retention (EV ratio up to 1.12) but leaves total illumination below 46-53% of ground truth.","lead":"The authors train a conditional generative model, AllSky, on physically captured HDR sky images to produce environment maps with better-preserved dynamic range. The paper introduces exposure-aware cascade losses and learned LDR-to-EDR tonemappers, arguing that prior sky models lose the wide dynamic range needed for image-based lighting.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The bundled claim of 'greater retention of dynamic range and illumination' is contradicted by the paper's own Table 1 vs Table 3 illumination ratios, and the dynamic-range half rests on a nonstandard, unvalidated EV metric.","rationale":"I read the paper in good faith: the proposed Selective/Cascade losses and the ldr2EDR and latent2EDR heads are plausible engineering ideas, and the ablations show internal gains in the reported EV ratio. However, the central contribution as stated in §1 and the abstract promises both dynamic range and illumination retention relative to previous DNN sky-models. The paper's own quantitative definition of illumination is integrated illumination (eq. 6), and Table 1 gives DeepClouds 0.59/0.62 while Table 3 gives AllSky at most 0.53. That is a direct internal contradiction, not a matter of contested metrics. The dynamic-range half is also insecure because EV is defined as log2(max−min), which is not standard dynamic range and is not scale-invariant; but I do not need to rely on that to reject the headline claim. The reader's REJECT is appropriate; my concrete test would pin the most decisive issue. I also note the paper's honest limitation statements (cloud segmentation variability, weak visual losses, HDRDB ghosting) count as credit, but they do not repair the central contradictory claim.","tokens_in":12892,"tokens_out":10080,"duration_ms":94713,"concrete_test":"Run AllSky's best configuration and the DeepClouds baseline on the same held-out HDRDB test subset with identical exposure normalization, computing the integrated-illumination ratio defined in §4.1.2. If AllSky's ratio remains below DeepClouds' 0.59/0.62, the illumination component of the central claim is false; if the reported gap shrinks or reverses under matched normalization, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (§1 and abstract) is that AllSky produces environment maps with 'greater retention of dynamic range and illumination' than previous DNN sky-models. The illumination part is refuted by the paper's own tables: Table 1 reports integrated-illumination ratios of 0.59 for DeepClouds and 0.62 for DeepClouds without the clear-sky sun, while the best AllSky configuration in Table 3 reaches only 0.53. These ratios use the metric the paper itself defines in §4.1.2 as global illumination, so no re-reading of 'retention of illumination' can reconcile the claim with the measurements unless the tables use different test splits or exposure normalization, which the text does not document. This is load-bearing because the abstract and introduction promise both EDR retention and illumination retention; if the illumination half fails, the headline contribution is reduced to a dynamic-range claim, which is itself undermined by the nonstandard EV = log2(max−min) metric in §3.1. The paper's own limitation discussion acknowledges weak cloud texture, but the illumination shortfall is more serious: it contradicts the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AllSky, a conditional generative sky-model trained on the Laval HDR Sky database, which takes user-provided sun-position and cloud-formation label maps and outputs HDR environment maps. The method introduces a custom μ-lawLog2 tonemapper, exposure-bracketed 'Cascade' losses, class-segmented selective losses, and ANN-based LDR-to-EDR (ldr2EDR and latent2EDR) heads. The central claim, stated in Section 1 and the abstract, is that AllSky produces environment maps with greater retention of dynamic range and illumination than prior DNN sky-models. The paper evaluates this through an EV ratio metric defined as log2(|I|max - |I|min), an integrated illumination ratio, and a set of visual metrics.","tokens_in":13134,"tokens_out":3833,"duration_ms":36808,"significance":"If the central claim were correctly demonstrated, the work would be a useful step toward controllable, physically-plausible HDR sky environment maps for image-based lighting. The paper also raises a legitimate concern that standard LDR-space losses are insensitive to the high dynamic range of solar regions. However, the significance is heavily undermined by the evaluation methodology: the custom EV metric is nonstandard and unvalidated, the illumination-retention claim is contradicted by the paper's own tables, and the training objective is constructed so that the headline metric is directly optimized, creating a circularity concern. The paper does not ship code or pre-trained models, and visual results are acknowledged to lose cloud texture, so the practical contribution is currently limited.","major_comments":[{"comment":"The dynamic-range metric is defined as EV = log2(|I|max - |I|min), i.e., the logarithm of an intensity difference, rather than the conventional definition EV = log2(max/min). This definition is scale-dependent and does not measure the number of stops between the dimmest and brightest parts of the image in any standard sense. All claims about 'Extended Dynamic Range' retention in Tables 2 and 3 are computed with this metric. No validation is provided that this quantity tracks usable dynamic range for image-based lighting, and no comparison to standard metrics (e.g., the ratio of the 99th to 1st percentile radiance, or the max/min ratio) is given. If the formula is unrepresentative, the headline quantitative results collapse. This is load-bearing and needs to be fixed or replaced.","section":"Section 3.1 (also Introduction)"},{"comment":"The introduction claims that AllSky demonstrates 'greater retention of dynamic range and illumination' (Section 1, final paragraph). The illumination part is refuted by the paper's own measurements. Using the same integrated illumination ratio defined in Section 4.1.2 (Eq. 6), the best AllSky configuration in Table 3 (ldr2EDR-9 HDR) reaches an illumination ratio of 0.53, while Table 1 reports DeepClouds at 0.59 and DeepClouds w/o Clear-Sky at 0.62. Even ignoring the outlier DeepClouds w/ Sun (3e31, caused by the sun pass-through mechanism), AllSky does not beat the baseline on illumination. Since the abstract and introduction promise both dynamic-range and illumination retention, this internal contradiction materially undermines the central contribution. The authors should either revise the claim or provide a fair comparison that reconciles these numbers.","section":"Section 4.3.4, Table 3 vs. Section 4.3.1, Table 1"},{"comment":"The Cascade losses explicitly segment the image by exposure brackets, with masks M = 2^{i-1} ≤ I ≤ 2^i and re-exposure by 2^{-i}, and penalize errors per bracket. The headline evaluation metric, EV ratio, is computed from the span of exposure values in the output. Training with these losses directly optimizes the quantity being measured: a model that matches the per-bracket intensity statistics will trivially improve the max-minus-min span. The reported EV gains therefore do not independently validate dynamic-range retention; they may simply reflect the training objective. To make the evaluation non-circular, the paper should either use an independent dynamic-range metric (e.g., percentile-based stops or a downstream IBL validation) or explicitly show that the metric improves on test data in a way that is not a direct artifact of the loss construction.","section":"Section 3.1.1, Eqs. (2)-(4) and Section 4.2.5"}],"minor_comments":[{"comment":"The abstract promises 'improved retention of the Extended Dynamic Range (EDR) of the sky', while Section 1 promises 'greater retention of dynamic range and illumination'. The discrepancy matters because the illumination claim is not supported by the tables; the authors should align these statements.","section":"Abstract and Section 1"},{"comment":"The paper states that the cloud segmentation via color-ratio thresholding is 'robust but variable under different lighting and seasonality' and later acknowledges inconsistency (Section 5). Given that the segmentation directly supervises the class-aware losses, this limitation should be discussed more thoroughly, e.g., with an estimate of segmentation error on the test set.","section":"Section 4.1.3"},{"comment":"The discussion concedes that 'cloud textures are essentially lost for all DNN sky-models', which is visually evident in Figures 8, 10 and 11. This is a significant caveat to any claim of 'photorealistic' output and should be stated more prominently, not only as a closing limitation.","section":"Section 5"},{"comment":"The table reports FID for AllSky but Table 1 does not report FID for DeepClouds, making the visual-quality comparison incomplete. Also, several AllSky configurations have FID values above 100 (latent2EDR-9 HDR), which are much worse than typical trained generative models; this deserves commentary.","section":"Table 3"},{"comment":"The exposure bracket masks are defined as M = 2^{i-1} ≤ I ≤ 2^i. For i=0 this includes negative intensities if I can be negative; the text says exposure is floored such that 2^{-1}=0, but the notation is not self-explanatory. Clarify the treatment of negative or zero intensities.","section":"Equations (2)-(4)"},{"comment":"The Text2Light evaluation uses only 80 image pairs and matches them via ORB keypoints plus a scale-invariant loss. The paper should state the matching accuracy, since misalignment could affect the reported L1 and EV values. It would also help to report the number of unique prompts versus duplicated generated images (the text mentions 63 duplicates but does not give a clear breakdown).","section":"Section 4.2.4"}],"recommendation":"reject","confidential_remarks":"The paper has a potentially relevant idea (exposure-bracketed losses for HDR sky generation), but the evaluation is not sound. The EV metric is nonstandard and unvalidated, the illumination-retention claim is contradicted by the paper's own Table 1 and Table 3, and the training loss directly optimizes the headline EV metric, making the reported gains circular. These are load-bearing problems that would require re-deriving the evaluation and likely re-running all experiments, which is not a minor fix. I cannot recommend publication in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper has a genuinely useful idea buried under a misleading headline. The authors propose exposure-segmented cascade losses and class-selective losses, plus a learned LDR-to-EDR expansion head (ldr2EDR / latent2EDR) that lets a U-Net train in tonemapped space without losing the sun's brightness. That is a real contribution – I haven't seen those specific components in the cited DeepClouds, SkyGAN, LM-GAN, or Text2Light. The paper also does a good job demonstrating that standard visual metrics (LPIPS, FID) are insensitive to dynamic range truncation, which is an important point for the IBL community. The evaluation of Text2Light's duplicates and seam artifacts is a useful side result.\n\nThe problems are in the evidence. The abstract only claims \"improved retention of EDR,\" but the introduction claims \"greater retention of dynamic range and illumination.\" The illumination half is not supported by their own numbers. Table 3 gives AllSky integrated-illumination ratios of 0.36–0.53, while Table 1 shows DeepClouds at 0.59 and DeepClouds without clear-sky at 0.62. The tables may use different training sets and resolutions, but the claim is unqualified and the paper does not discuss this discrepancy. That is a load-bearing flaw.\n\nThe dynamic range metric is also trouble. They define EV = log2(|I|max − |I|min), which is not a ratio and does not measure the number of stops between max and min. Standard EV/DR is log2(max/min). As defined, a scene with a bright sun and near-black sky could have the same EV as a uniform scene if the absolute spread is the same. And since their cascade losses explicitly supervise per exposure bracket, the EV ratio they report is essentially what the loss is optimizing – some circularity is unavoidable. They give no validation that this metric tracks usable dynamic range for IBL.\n\nThere is also no code, no error bars, and the AllSky experiments are at 64x64 on 5,000 images, far from the scale of the DeepClouds reproduction. The paper honestly acknowledges cloud-texture loss and segmentation inconsistencies, but the metric and claim issues are more serious than those limitations.\n\nWho is this for? People working on sky models, HDR generation, and image-based lighting. The losses and the ldr2EDR idea are worth knowing about, even if the evaluation is not convincing. I would not desk-reject it: it has a novel technique and a valid problem. A careful reviewer could help the authors reframe the claims, fix the metric, and run apples-to-apples comparisons. Needs major revision, but there is something here.\n\nRecommendation: send to peer review, with a strong request for code, error bars, and a metric justification.","headline":"A promising set of HDR-aware losses and a learned LDR-to-EDR head, but the illumination claim is contradicted by the paper's own tables and the EV metric is nonstandard.","tokens_in":13703,"tokens_out":4580,"would_cite":false,"duration_ms":41557,"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":"AllSky, a DNN trained on physically captured HDR skies, generates environment maps that retain more of the sun's dynamic range than prior deep sky-models, using exposure-bracketed class-aware losses and a learned LDR-to-EDR booster.","keywords":["sky-model","sun","clouds","high dynamic range","environment maps","image-based lighting","extended dynamic range","HDR image synthesis"],"falsifier":"Recompute the headline comparisons with EV defined as log2(max/min) and with integrated illumination measured after exposure equalization; if AllSky no longer beats DeepClouds and Text2Light on those recomputed ratios, the central claim that it retains more extended dynamic range fails.","tokens_in":12641,"feed_emoji":"☀️","tokens_out":8522,"duration_ms":68348,"temperature":0.7,"pith_summary":"The paper argues that previous deep-learning sky-models, despite improving visual quality, quietly lose the wide dynamic range of real outdoor light, so generated environment maps relight scenes with wrong tones and shadows. It proposes AllSky, a conditional generator trained on tens of thousands of physically captured HDR skies, paired with two mechanisms: losses that weight errors by exposure bracket and by region (sun, sky, cloud) and a learned inverse-tone-mapping network that expands a low-dynamic-range output back toward 14 stops. In comparisons, AllSky reports higher ratios of generated-to-ground-truth dynamic range and integrated illumination than a cloud-augmenting baseline and a text-driven HDR panorama model. The paper's deeper point is that dynamic range, not average pixel error, is what makes an environment map usable for image-based lighting.","feed_headline":"AllSky sky model preserves more of the sun's dynamic range","feed_subtitle":"Sun-aware losses and a learned LDR-to-HDR booster make generated outdoor lighting retain more of the physical scene's range.","key_machinery":"The load-bearing mechanism is a two-stage training setup. First, a hybrid tone-mapper, μ-lawLog2, compresses HDR intensities to a range near −1 to 1, preserving low-exposure cloud texture while aggressively compressing saturated solar regions. Second, the loss function is split by exposure bracket (Cascade loss) and by semantic class (Selective loss), so the small, ultra-bright sun region is not numerically drowned out by the large dim sky. Third, an ANN head called ldr2EDR (or latent2EDR, operating on the U-Net's latent space) learns the inverse tone-mapping from the network's compressed output back to linear EDR values. The paper evaluates with the EV ratio and the integrated illumination ratio rather than relying on L1, L2, or perceptual metrics alone.","core_discovery":"The central claim is that a DNN sky-model can learn weathered skies directly from physically captured HDR imagery and, with the right loss structure and a learned decompression head, retain a substantially larger fraction of the scene's extended dynamic range than prior DNN models. The paper shows that conventional L1 and L2 training in tonemapped space collapses the solar region's intensity, and that adding Selective and Cascade losses—which split the image into exposure bands and semantic classes—plus an ANN LDR2EDR booster recovers a larger share of ground-truth EV and integrated illumination. Reported results put AllSky's generated-to-ground-truth EV ratios near 0.85–1.12 and illumination ratios near 0.5, versus about 0.36 EV for the DeepClouds baseline and a Text2Light HDR overshoot of 2.42 EV and 3.55 times illumination, on the paper's own metrics.","pith_inferences":["If the dynamic-range comparisons are recomputed with the standard log-ratio definition of EV, log2(max/min), the gap between AllSky and its baselines may shrink, because the paper's formula log2(|I|max − |I|min) rewards intensity differences rather than contrast ratios.","The LDR2EDR head is color-agnostic and trained on a fixed dataset, so it could likely be transferred as a post-processor to any generator whose output is tonemapped LDR, though its behavior on non-sky content is untested.","A learned cloud segmentation, or labels derived from a weather model, would probably make the Selective losses more effective than the color-ratio threshold, since the paper itself notes that segmentation is inconsistent across seasons.","For image-based lighting, the practical test is not the EV ratio but rendered appearance; a direct comparison of shadows and specular highlights at matched exposures would settle whether the extra dynamic range actually changes the relit scene."],"forward_implications":["AllSky outputs, conditioned on a user-drawn sun and cloud label, retain a larger fraction of ground-truth EV and integrated illumination than DeepClouds and Text2Light, so rendered scenes should show tones and shadows closer to the physical capture.","Cascade and Selective losses can be dropped into other HDRI generation pipelines, since they only replace the loss function rather than the architecture.","The learned LDR2EDR head can extend the dynamic range of an already-trained U-Net backbone with negligible change to visual-quality metrics such as LPIPS and FID.","Dynamic range and integrated illumination should be reported alongside L1/L2 and perceptual scores when evaluating any sky-model, because conventional metrics are shown to be insensitive to the clipping that changes image-based lighting.","The parametric clear-sky sun aggregation used by DeepClouds is what gives it dynamic range, but the paper finds it produces a tiny, visually unappealing solar disc that can pierce clouds, a failure AllSky avoids by learning the sun from data."],"supporting_citations":[{"why":"Supplies the 34K+ physically captured HDR sky images used for training, validation, and testing AllSky.","marker":"[19]"},{"why":"DeepClouds is the cloud-augmenting DNN baseline whose dynamic range and illumination AllSky is compared against.","marker":"[33]"},{"why":"Text2Light is the text-conditioned HDR panorama baseline that AllSky is compared against for EDR retention.","marker":"[5]"},{"why":"UNetFixUp is the U-Net backbone whose output layer AllSky replaces with the ldr2EDR and latent2EDR heads.","marker":"[11]"},{"why":"Provides the color-ratio cloud segmentation used to build the class masks for the Selective losses.","marker":"[7]"},{"why":"Supplies the solar ephemeris positions used to place the sun-region label in the training masks.","marker":"[36]"},{"why":"Defines the direct HDR sun and sky capture method used to build the physical ground-truth database.","marker":"[37]"},{"why":"The parametric clear-sky model that the DeepClouds baseline augments with clouds, providing the sun-pass-through mechanism analyzed in the comparison.","marker":"[17]"}],"fun_headline_variants":["AllSky keeps more of the sun's dynamic range in sky relighting","Sky model preserves extended dynamic range from real HDR captures","DNN sky model retains solar EV with selective losses and booster","AllSky: learning weathered skies with improved EDR retention","Sky relighting gains from preserving sun's dynamic range"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the paper's EV formula, log2(|I|max − |I|min), is a meaningful measure of the dynamic range that matters for image-based lighting, since all headline comparisons use it and the standard log-ratio definition would give different numbers.","fun_headline_variants_meta":{"raw":{"variants":["AllSky keeps more of the sun's dynamic range in sky relighting","Sky model preserves extended dynamic range from real HDR captures","DNN sky model retains solar EV with selective losses and booster","AllSky: learning weathered skies with improved EDR retention","Sky relighting gains from preserving sun's dynamic range"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1364,"prompt_tokens":948,"completion_tokens":416,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":332}},"tokens_in":564,"tokens_out":416,"duration_ms":4525,"temperature":1.0,"reasoning_tokens":332,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:29:23.027859+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the headline comparisons with EV defined as log2(max/min) and with integrated illumination measured after exposure equalization; if AllSky no longer beats DeepClouds and Text2Light on those recomputed ratios, the central claim that it retains more extended dynamic range fails.","supporting_citations":[{"cited_title":"Lalonde, L.-P","cited_arxiv_id":null,"evidence_quote":"Supplies the 34K+ physically captured HDR sky images used for training, validation, and testing AllSky."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DeepClouds is the cloud-augmenting DNN baseline whose dynamic range and illumination AllSky is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Text2Light is the text-conditioned HDR panorama baseline that AllSky is compared against for EDR retention."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"UNetFixUp is the U-Net backbone whose output layer AllSky replaces with the ldr2EDR and latent2EDR heads."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the color-ratio cloud segmentation used to build the class masks for the Selective losses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the solar ephemeris positions used to place the sun-region label in the training masks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the direct HDR sun and sky capture method used to build the physical ground-truth database."}],"review_version":1}