{"id":"4ae178a2-5725-43f1-9725-f4ac295a9450","arxiv_id":"2508.18531","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 3D building generator that combines satellite imagery with coarse geometric priors via cosine interpolation between noise and prior latents, trained on a new 50K building dataset.","lead":"The paper introduces SatSkylines, a system that generates 3D building models from overhead satellite photos and simple geometric hints such as boxes or building footprints. It also contributes Skylines-50K, a dataset of over 50,000 game-derived 3D buildings, and claims fast, controllable generation for large-scale city modeling.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quantitative results use ground-truth-derived priors, not real OSM priors; without a realistic prior-noise test, the 0.9381 IoU does not support the claimed real-world generalization.","rationale":"The reader's weakest assumption is also my principal concern. Every headline number in Tables 1 and 2 is conditioned on priors that are exact derivatives of the target asset. The abstract's claim of strong generalization can only be true if the model tolerates the noise, misalignment, and incompleteness of actual OpenStreetMap priors. The paper provides no quantitative evidence on that shifted distribution; the real-world section is qualitative, and Supplementary Section 2 explicitly limits expectations. This is not an internal inconsistency of the architecture — the interpolation mechanism and dataset construction are plausible — but it is the point where the central claim is least secure. Other issues, such as the 20-asset CLAY comparison, missing error bars, and licensing questions, are secondary and addressable. The proposed noisy-prior test is cheap and directly measures the gap between benchmark and deployment. If the model degrades gracefully under realistic prior noise, the conditional should be lifted; if it does not, the headline numbers overstate in-the-wild performance. Therefore I keep the reader's CONDITIONAL verdict.","tokens_in":13227,"tokens_out":6716,"duration_ms":66163,"concrete_test":"On the existing 500-asset test set, create a noisy-prior condition: for each asset, take the LOD1 prior, perturb the footprint with a random 2 m translation and 5° rotation, replace the polygon with its convex hull, and add ±5% height jitter; regenerate with the released model at λ=0.5 and recompute IoU, CD, and F-score. If the mean IoU drops by more than 0.1 relative to the clean-prior condition, the clean benchmark materially overstates robustness to real OSM priors. If the drop is negligible, the generalization concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 states that because the dataset has no OpenStreetMap data, the authors generate coarse geometric priors from the ground-truth assets themselves: LOD 1 is the target building's exact footprint extruded to its exact height, and LOD 2 uses two exact cross-sections. Section 3.3 then claims the real-world pipeline uses OSM footprints and height attributes. These distributions differ: OSM footprints are simplified, sometimes offset by meters, and heights are noisy or missing. Training never exposes the model to such misalignment. Therefore the headline IoU of 0.9381 (Table 1) is measured under perfect-prior conditions and is an upper bound, not the in-the-wild operating point. The paper provides no quantitative real-world 3D evaluation; real-world results are qualitative (Figs. 5 and 10), and Supplementary Section 2 concedes outputs are not expected to be identical to ground truth. The central claim of strong generalization thus rests on an untested distribution shift between exact synthetic priors and noisy OSM priors.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SatSkylines, a method for generating detailed 3D building models from top-down satellite imagery and coarse geometric priors such as cuboids or building footprints. The key idea is to encode a coarse geometry prior with a sparse-structure VAE, apply channel-wise latent normalization, and then cosine-interpolate the normalized latent code with Gaussian noise before feeding it to a rectified-flow transformer that generates geometry and appearance. The method is built on the Trellis architecture and is trained on a newly curated dataset, Skylines-50K, consisting of over 50,000 building assets harvested from the game Cities: Skylines. The paper also presents an end-to-end pipeline that, given a GPS bounding box, extracts OpenStreetMap footprints and heights as priors, enhances satellite imagery with a GPT-based model, and generates 3D buildings. Quantitative experiments on a 500-instance test set report large gains over Trellis and CLAY (IoU 0.9381 vs 0.4415 for Trellis), together with ablations and qualitative real-world results.","tokens_in":13389,"tokens_out":5735,"duration_ms":53251,"significance":"If the reported results are robust, the paper makes a useful contribution: a simple, computationally cheap conditioning mechanism (cosine interpolation with latent normalization) that enables geometric control in image-conditioned 3D generation, and a large, diverse, textured 3D building dataset that could support further research. The method also demonstrates fast inference (~15 s per building), which is valuable for city-scale applications. The paper is commendable for including a finetuned Trellis baseline in the ablations and for clearly describing the dataset curation. However, the evidence for real-world generalization is currently qualitative, and the quantitative evaluation is conducted under conditions that systematically differ from the real-world pipeline. The central claim of strong generalization therefore requires additional experimental support.","major_comments":[{"comment":"The quantitative evaluation uses coarse geometric priors generated from the ground-truth assets themselves: Section 3.2 states that LOD 1 is the target building's exact footprint extruded to its exact height, and LOD 2 uses two exact cross-sections. In contrast, the real-world pipeline in Section 3.3 uses OpenStreetMap footprints and height attributes, which are typically simplified, sometimes offset by meters, and often noisy or incomplete. The model is never trained or evaluated on such misaligned priors, so the reported IoU of 0.9381 in Table 1 is an upper bound measured under perfect-prior conditions and does not by itself substantiate the claimed strong generalization to real-world data. I request an additional experiment: either evaluate the method on real OSM-derived priors (with reference 3D models from LiDAR or photogrammetry), or simulate realistic prior noise on the existing test set (e.g., random footprint simplification, translation offsets, height quantization) and report the resulting metric degradation.","section":"§3.2, §3.3, Table 1"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any of the quantitative results. This is particularly problematic for the CLAY comparison, which is carried out on only 20 assets (Table 1), and for the ablations where the differences are small, such as the 0.003 Chamfer Distance difference between SatSkylines and 'wo ss latent norm' in Table 3, and the LOD 0 versus LOD 1 IoU difference in Table 2. Without variance estimates, it is unclear whether the reported improvements, especially those used to justify the latent normalization and the ordering of LOD levels, are meaningful. The authors should provide multiple runs, bootstrap confidence intervals, or at least error bars over test-set splits.","section":"§4.1, Tables 1–3"},{"comment":"The image-condition distribution used in evaluation is also aligned with the ground truth: the top-down images for Skylines-50K are rendered from the same 3D assets, whereas the real-world pipeline uses satellite imagery from Google Maps or Mapbox, which differs in resolution, scale, lighting, and imaging artifacts, and is further post-processed by 'gpt-image-1'. The paper provides only qualitative real-world results and no quantitative assessment of this domain gap. I recommend a quantitative real-world evaluation on a small set of buildings with available reference geometry, or at minimum a clear statement that the quantitative results apply to the synthetic rendered setting and that real-world performance is likely lower.","section":"§3.3, §4.1, Fig. 7"}],"minor_comments":[{"comment":"The word 'Condtional' in the first sentence should be corrected to 'Conditional'.","section":"§1"},{"comment":"The phrase 'we empoly Chamfer Distance' contains a typo; 'empoly' should be 'employ'.","section":"§4.1"},{"comment":"The caption contains the typo 'Geormtric Priors'; it should be 'Geometric Priors'.","section":"Table 2 caption"},{"comment":"The paper does not specify how the 20-instance sub-test set was selected for the CLAY comparison. For reproducibility, please describe the selection process (e.g., random or stratified).","section":"§4.1"},{"comment":"The prompt used for 'gpt-image-1' appears only in the supplementary material. Since the pipeline depends on this prompt, consider providing the full prompt in the main text or referencing it more prominently, as the model output is not reproducible without it.","section":"§3.3, Supplementary Fig. 9"},{"comment":"The statement that the dataset 'will be released' appears only in the supplementary material; it should be stated in the main text, as it is an important contribution.","section":"Supplementary §3"},{"comment":"To avoid confusion, the caption of Table 3 should clarify that 'Trellis w finetune on Skylines-50K' is the image-only Trellis model finetuned on the training split of Skylines-50K, in contrast to the un-finetuned Trellis row in Table 1.","section":"Tables 1 and 3"},{"comment":"The paper reports IoU, CD, and F-score on 'generated sparse structure voxels', but does not state the voxel resolution used for evaluation. Please specify the grid resolution to enable fair comparison with future work.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a clean and simple method, and the dataset is a valuable resource. The main weakness is the mismatch between the evaluation protocol and the claimed real-world generalization: the quantitative numbers are obtained with ground-truth-derived priors and rendered images, while the real-world pipeline uses noisy OSM data and different satellite imagery. If the authors can add experiments with realistic prior noise or a small real-world quantitative evaluation, the central claim would be much better supported. The lack of error bars is also a concern, especially for the 20-instance CLAY comparison and the small ablation differences. I recommend major revision rather than rejection because the issues appear addressable within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead SatSkylines. The useful core is the conditioning mechanism: they take a coarse geometric prior, encode it with Trellis's sparse-structure VAE, normalize the latent channel-wise, then cosine-interpolate between that normalized prior latent and Gaussian noise, with a strength parameter λ. That is a clean answer to the one-to-many problem in coarse-to-fine generation, and it adds no inference cost. The Skylines-50K dataset (50k textured buildings from the Cities: Skylines workshop) is a real resource with genuinely global diversity. The paper is clearly written, and the ablations are honest — they even finetune Trellis on their data, which narrows the baseline gap.\n\nThe soft spot is not the method; it is the evidence for real-world generalization. Section 3.2 says they generate the coarse priors from the ground-truth assets themselves because the dataset has no OSM data. LOD 1 is the exact footprint extruded to the exact height; LOD 2 uses exact cross-sections. The real-world pipeline (Sec 3.3) uses OSM footprints and height attributes, which are simplified, offset, noisy, sometimes missing. Those are different distributions. The Table 1 IoU of 0.9381 vs Trellis's 0.4415 is measured under perfect-prior conditions. It shows what the method can do when the prior is right; it does not measure what happens when the prior has typical OSM errors. The real-world results are qualitative only, and the supplementary concedes outputs are not expected to match ground truth. So the claim of \"strong generalization\" is an extrapolation, not a measured fact.\n\nMinor issues: no error bars, CLAY is evaluated on only 20 assets, the gpt-image-1 super-resolution step is a closed API, and dataset licensing is unaddressed. None of these sink the paper; the core idea is sound and reproducible enough for others to build on.\n\nBottom line: if you work on 3D building generation or conditional 3D generation, this is worth reading and citing. It deserves peer review, with a request for a realistic prior-noise evaluation — perturbing the exact priors with OSM-like displacement, or a small labeled real-world set with LiDAR/GIS ground truth. I would send it out.","headline":"A genuinely new conditioning mechanism and a useful dataset, but the headline numbers are measured with ground-truth-derived priors, so real-world generalization is asserted rather than shown.","tokens_in":13939,"tokens_out":1875,"would_cite":true,"duration_ms":17344,"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":"Satellite views plus cuboid priors produce detailed 3D buildings","keywords":["3D building generation","satellite imagery","coarse geometric priors","flow transformer","cosine interpolation","Skylines-50K dataset","rectified flow","top-down view"],"falsifier":"Take a set of GPS coordinates covered by OpenStreetMap, run the full pipeline to get OSM-derived priors, then compare the generated geometry against airborne LiDAR scans of the same buildings: if the median voxel IoU against the LiDAR ground truth falls well below the 0.938 reported on clean priors (say below 0.7), then the clean-prior assumption is the limiting factor and the headline accuracy does not carry over to in-the-wild data.","tokens_in":1713,"feed_emoji":"🏢","tokens_out":5542,"duration_ms":89766,"temperature":0.7,"pith_summary":"The paper proposes SatSkylines, a generative model that turns a top-down satellite image and a very coarse geometric prior—like a simple box or a footprint-based shape—into a detailed, textured 3D building. The central trick is to feed the model an interpolation between the latent code of the coarse prior and random Gaussian noise, so the model learns a one-to-many mapping from simple priors to many plausible detailed buildings. On a 500-building test set the method reaches a geometry IoU of 0.938, far above the 0.442 scored by Trellis with image-only input, and does so in about 15 seconds per building. If the result holds up in the wild, it would let city-scale 3D modeling be driven by globally available satellite imagery and OpenStreetMap footprints, without needing detailed voxel models or ground-level photos.","feed_headline":"Satellite views plus cuboid priors produce detailed 3D buildings","feed_subtitle":"0.94 IoU on test buildings, with 15-second generation per asset.","key_machinery":"The load-bearing mechanism is cosine geometric interpolation between a channel-wise normalized SS-VAE latent of the coarse geometric prior and Gaussian noise, done before the sparse-structure Rectified Flow transformer. The normalization forces the prior latent to look like noise (roughly $\\mathcal{N}(0,1)$), and the identity $\\cos^2(\\lambda\\pi/2)+\\sin^2(\\lambda\\pi/2)=1$ guarantees the interpolated input stays Gaussian, so the model is trained on a smooth path from 'mostly prior' to 'pure noise' and at inference λ acts as a geometric-guidance dial.","core_discovery":"SatSkylines establishes that a single sparse-structure flow transformer, trained on a large set of stylized building assets, can generate detailed and appearance-consistent 3D buildings conditioned on a pure top-down satellite image and a coarse geometric prior as simple as one or two stacked cuboids. The authors achieve this by encoding the coarse prior into the SS-VAE latent space, normalizing that latent channel-wise so it matches the statistics of Gaussian noise, and then cosine-interpolating between the normalized prior latent and pure noise with a strength parameter λ. This keeps the input distribution Gaussian because $\\cos^2(\\lambda\\pi/2)+\\sin^2(\\lambda\\pi/2)=1$, lets the same training recipe as Trellis apply, and gives the user a dial between fidelity to the prior and creative freedom. In their experiments the model reaches 0.9381 IoU and 0.8684 F-score on a 500-instance test set, outperforming image-only Trellis (0.4415 IoU) and voxel-hungry CLAY (0.6859 IoU on a 20-instance subset), with roughly 15-second per-building inference.","pith_inferences":["The cosine interpolation's Gaussian-preserving property is not specific to buildings: the same 'normalize the condition latent, then interpolate with noise' recipe could make any image-plus-shape-conditioned flow model accept much coarser geometric controls, provided the condition encoder's output distribution is approximately Gaussian.","A testable extension would be to vary λ per building in a real-world city reconstruction, letting the user trade geometric fidelity against hallucinated detail; for landmarks such as the Arc de Triomphe a strong prior matters, while for generic infill a lower λ may give more plausible variety.","The paper's real-world evidence is qualitative only; a quantitative study using airborne LiDAR or photogrammetric meshes as ground truth for a few city blocks would be the natural next step to verify whether the 0.94-level IoU transfers to noisy OSM-derived priors.","If the method generalizes, it implies that detailed 3D city models could be generated for any location with satellite coverage and OSM data, potentially changing the cost curve for digital twins, gaming, and urban planning."],"forward_implications":["With only a cuboid prior (LOD 0), SatSkylines still achieves 0.9515 IoU on the test set, showing the method degrades gracefully as priors get coarser.","Richer priors monotonically improve reconstruction fidelity (Chamfer Distance drops from 0.0222 at LOD 0 to 0.0141 at LOD 2), so the same model benefits from any extra footprint or height information OpenStreetMap can provide.","Because the interpolation is applied at inference only and adds no parameters, the approach preserves Trellis-like speed, generating a textured building in about 15 seconds rather than the minutes required by CLAY or Rodin.","The Skylines-50K dataset (50,673 hand-authored, textured building assets with PBR materials) can serve as a training resource for other satellite-to-3D building generation models.","The end-to-end GPS pipeline (OpenStreetMap footprints plus heights to coarse prior, satellite image retrieval to GPT-based super-resolution to generation) means the method can be applied to arbitrary coordinates without manual voxel editing."],"supporting_citations":[{"why":"Supplies the base architecture (SS and SLat flow transformers, image conditioning via cross-attention) that SatSkylines builds on, and serves as the primary image-only baseline.","marker":"[55]"},{"why":"Demonstrates direct coarse-to-fine geometry flow, providing the one-to-many mapping contrast that motivated the interpolation approach.","marker":"[10]"},{"why":"Is the voxel-conditioned baseline that requires detailed priors and slow inference, compared on the 20-instance subset.","marker":"[65]"},{"why":"Defines the Rectified Flow training objective used for both the SS and SLat transformers.","marker":"[39]"},{"why":"Provides the DINOv2 visual features used for image conditioning in the cross-attention layers, following Trellis' design.","marker":"[41]"}],"fun_headline_variants":["Cuboid priors + satellite views yield detailed 3D buildings","Satellite image + a box becomes fine 3D building","One cuboid prior plus satellite view makes detailed 3D","Coarse priors plus satellite imagery equal detailed 3D building models","From simple cuboids to detailed 3D buildings with satellite images"],"cache_read_input_tokens":16128,"weakest_assumption_plain":"The reported accuracy is measured on coarse priors generated from the ground-truth building shapes themselves, not from noisy or incomplete real-world OpenStreetMap data, so the in-the-wild performance depends on real priors being similar enough to those clean priors.","fun_headline_variants_meta":{"raw":{"variants":["Cuboid priors + satellite views yield detailed 3D buildings","Satellite image + a box becomes fine 3D building","One cuboid prior plus satellite view makes detailed 3D","Coarse priors plus satellite imagery equal detailed 3D building models","From simple cuboids to detailed 3D buildings with satellite images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001102,"raw_usage":{"total_tokens":4591,"prompt_tokens":934,"completion_tokens":3657,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":3566}},"tokens_in":550,"tokens_out":3657,"duration_ms":25548,"temperature":1.0,"reasoning_tokens":3566,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:56:46.658465+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of GPS coordinates covered by OpenStreetMap, run the full pipeline to get OSM-derived priors, then compare the generated geometry against airborne LiDAR scans of the same buildings: if the median voxel IoU against the LiDAR ground truth falls well below the 0.938 reported on clean priors (say below 0.7), then the clean-prior assumption is the limiting factor and the headline accuracy does not carry over to in-the-wild data.","supporting_citations":[{"cited_title":"Structured 3d latents for scalable and versatile 3d gen- eration","cited_arxiv_id":null,"evidence_quote":"Supplies the base architecture (SS and SLat flow transformers, image conditioning via cross-attention) that SatSkylines builds on, and serves as the primary image-only baseline."},{"cited_title":"Clay: A controllable large-scale generative model for creat- ing high-quality 3d assets","cited_arxiv_id":null,"evidence_quote":"Is the voxel-conditioned baseline that requires detailed priors and slow inference, compared on the 20-instance subset."}],"review_version":2}