{"id":"f2bc2bf8-535a-4628-86b9-bc0215307694","arxiv_id":"2412.00157","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"AerialGo generates realistic ground-level city views from aerial images using a multi-view diffusion model and introduces a 3.45M-image synthetic urban dataset.","lead":"AerialGo uses a diffusion model to turn aerial photos of a city into street-level views, paired with a 3.45 million image synthetic dataset. The goal is cheaper, privacy-safe 3D city modeling, but the experiments only use synthetic cities.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central real-city claim rests on untested synthetic-to-real transfer; all experiments use UE5 renders, so ground-prior benefit on real aerial imagery is unsubstantiated.","rationale":"The reader's weakest_assumption — synthetic-to-real transfer — is the most load-bearing condition for the central claim. Even if the method perfectly improved ground-view reconstruction on UE5 data, the paper's stated purpose (real cities, privacy, scalability) would fail if the diffusion model and the point-cloud conditioning do not transfer to real aerial imagery. This is a missing empirical link rather than an internal inconsistency; the paper does not claim to test real data, but the framing in Sec. 1 and the abstract makes real-world applicability part of the claim. Secondary concerns include the mixed metrics in Table 2 (on the AerialGo City A block, SSIM and LPIPS degrade for 3DGS and CityGaussian when priors are added) and the lack of split aerial/ground metrics; these further reduce confidence but are not the single load-bearing issue. I agree with the reader's conditional verdict: the dataset and idea are plausible, but the claim needs a real-data demonstration, code/checkpoints, and separate ground-view metrics before acceptance. The proposed real-data test would settle whether transfer holds; if it passes, the central claim is credible, and if it fails, the paper should be reframed as synthetic-domain only.","tokens_in":12999,"tokens_out":4705,"duration_ms":40836,"concrete_test":"Use a small real urban block (e.g., ~0.1 km²) with drone or oblique aerial images and co-located ground-level images with known poses held out. Run the full pipeline: SfM/MVS on the aerial images to produce the conditioning point cloud, select N=3 nearest aerial references, generate ground views with the trained AerialGo model (or its released checkpoint), then reconstruct 3DGS and CityGaussian with and without these generated priors. Report ground-view PSNR, SSIM, LPIPS separately. If the with-priors condition does not improve all three metrics over aerial-only reconstruction, or if the generated ground views are visually inconsistent with real ground truth, the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Sec. 5.1 — that generated ground priors yield 'significantly improved reconstructions from the ground-level perspective' — is explicitly motivated (Sec. 1) as a privacy-conscious, scalable solution for real cities where ground-level capture is unavailable. However, every experiment in Sec. 5 trains and evaluates on Unreal Engine 5 renders (AerialGo dataset, MatrixCity). The diffusion model learns the distribution of UE5-rendered ground views from UE5 aerial views, with perfect camera poses, clean depth, and synthetic appearance (reflection glass, weather, foliage) that are not representative of real oblique aerial photogrammetry. The point-cloud conditioning is especially fragile: on real aerial data, SfM/MVS point clouds will be noisier, sparser, and contain holes, so the CLIP-encoded point-render condition used in cross-attention will be corrupted. The paper provides no real-world test, no domain-adaptation experiment, and no sensitivity analysis to conditioning noise, yet the abstract and intro claim a scalable real-city solution. The plausible mechanism is that the method relies on clean synthetic geometry and render statistics; if so, it fails in exactly the setting the paper is written to address. This is a missing empirical link, not a disagreement with consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AerialGo, a multi-view diffusion framework that generates ground-level city views from aerial images and uses them as priors for 3D reconstruction with 3DGS, CityGaussian, and GigaGS. It also introduces a large Unreal Engine 5 dataset of paired aerial and ground images covering 134 km² with camera, depth, and segmentation annotations. Experiments on the AerialGo and MatrixCity datasets compare reconstructions with and without generated priors, compare against existing novel-view-synthesis methods, and ablate the number of conditioning views and the point-render condition. The central claim is that generated ground views substantially improve the realism and structural coherence of ground-level urban reconstructions while avoiding direct ground-level data collection.","tokens_in":13191,"tokens_out":5135,"duration_ms":41014,"significance":"If fully supported, the result would be a useful step toward using generative priors to relieve the ground-data bottleneck in city-scale reconstruction, and the proposed 3.45M-image dataset would be a valuable community resource. The paper is also commendable for evaluating the priors on multiple reconstruction backbones and for including ablations of the conditioning design. There is no equation-level circularity: the diffusion model is trained on paired data and the reconstruction is evaluated against held-out ground truth. However, the evidence for the central claim is mixed: Table 2 shows performance degradation on AerialGo City A for two of three baselines, and all quantitative results are on synthetic UE5 renders, leaving the real-city applicability claimed in the introduction untested. The central method may be defensible, but the current manuscript needs additional experiments and more careful claims before it can be accepted.","major_comments":[{"comment":"The central claim of 'significantly improved reconstructions' is not consistently supported by the reported numbers. On AerialGo City A, adding priors lowers SSIM for 3DGS (0.686 to 0.626) and CityGaussian (0.674 to 0.654) and worsens LPIPS for 3DGS (0.328 to 0.397) and CityGaussian (0.350 to 0.387); only GigaGS improves across all three metrics. Because the table averages metrics over aerial and ground views, the paper should report ground-view-only metrics and per-block breakdowns, together with confidence intervals or significance tests, before the claim in Section 5.1 can be evaluated.","section":"Section 5.1, Table 2"},{"comment":"The 'with priors' condition changes three factors simultaneously: generated ground images are added, an LPIPS loss is introduced, and skybox modeling is added. The reported improvements therefore cannot be attributed to the generated ground-view priors alone. An ablation that isolates the generated priors (for example, generated images combined with the same LPIPS loss and skybox, versus no generated images under the same losses and skybox) is required to support the abstract's claim that AerialGo significantly enhances ground-level realism.","section":"Section 4.2 and Table 2"},{"comment":"All quantitative experiments are conducted on Unreal Engine 5 renders (AerialGo and MatrixCity), with perfect camera poses, clean depth, and synthetic appearance, yet the introduction and abstract frame the method as a solution for real cities where ground-level capture is unavailable. The point-cloud conditioning in Section 4.1 assumes that aerial point clouds reconstructed from images are accurate enough to render a useful CLIP condition; real SfM/MVS point clouds are typically noisier, sparser, and more incomplete. The paper needs at least one experiment on real oblique aerial data or an explicit sensitivity analysis to conditioning noise; without this, the real-city applicability claim is unsubstantiated.","section":"Sections 1, 4.1, and 5"},{"comment":"The comparison with LucidDreamer, ViewCraft, and MotionCtrl mentions 'necessary adaptations' without specifying what was adapted, which hyperparameters were used, or how the models were trained or fine-tuned. The reported PSNR values (7.98–10.46) are far below typical novel-view-synthesis results, so it is difficult to rule out configuration artifacts. Please provide implementation details or release code so that the comparison is verifiable and reproducible.","section":"Section 5.1, Table 3"}],"minor_comments":[{"comment":"The sentence 'such as ImageDream [44], LucidDreamer [7], ReconFusion [54], and ViewCraft [47]' appears to swap citation numbers: ViewCraft is reference [54] and ReconFusion is reference [47].","section":"Section 4.1 and References"},{"comment":"The text cites 'CityGaussian [17]', but reference [17] is the City3D paper, not the CityGaussian paper; the citation should be corrected to the appropriate entry.","section":"Section 5"},{"comment":"There are minor naming inconsistencies: 'MotionCrtl' should be 'MotionCtrl', and the paper alternates between 'GigaGs' and 'GigaGS'.","section":"Table 3 and Section 5.1"},{"comment":"The ablation table does not state which dataset it is run on or how the metrics are averaged over views, and the '0-view condition' row is not clearly distinguished from the 'w/o point render condition' row.","section":"Section 5.2, Table 4"},{"comment":"References [42] and [43] both refer to the same Mega-NeRF paper; the duplicate should be removed.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to require either a real-data demonstration or a substantial reframing of its claims before publication. The promised AerialGo dataset has no URL or access terms, so its availability should be verified. I also recommend asking the authors for ground-view-only metrics in Table 2, since the averaged numbers currently hide the degradation observed on AerialGo City A."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful dataset contribution and a reasonable method for generating ground views from aerial imagery, but the paper's central claim—that this enables scalable real-city reconstruction without ground data—is not supported by the evidence, which is entirely synthetic. The mixed numbers on their own benchmark make that gap more visible.\n\nWhat's genuinely new: the AerialGo dataset (3.45M image pairs, 134 km^2 of UE5 cities, with camera/depth/segmentation and weather/lighting variation) looks like a real resource if it's actually released. The method—multi-view diffusion based on ImageDream, with CLIP-encoded point-cloud renders as structural conditioning—is a sensible adaptation of existing ideas to a hard new regime. The reconstruction experiments are extensive: three 3DGS-family baselines, two datasets, and a comparison to three generative NVS methods where the gap is large and convincing.\n\nWhere it gets soft. The evidence for \"significantly improved reconstructions\" is mixed. On AerialGo City A, adding generated priors improves PSNR but worsens SSIM and LPIPS for two of the three baselines (e.g., 3DGS SSIM 0.686 to 0.626, LPIPS 0.328 to 0.397). On City B and MatrixCity the gains are consistent, but the claim as stated in Sec. 5.1 is too strong. More importantly, every experiment uses UE5 renders with clean poses, depth, and point clouds. The paper's motivation is real cities, but there is no real aerial test, no domain adaptation experiment, and no sensitivity analysis to noisy SfM/MVS point clouds—the exact condition that would break the point-cloud conditioning in practice. That is a missing empirical link, not a disagreement with consensus. Also, no comparison to InfiniCity or Sat2Scene, the closest prior works; no code or dataset release; no error bars; and several citation inconsistencies (ViewCraft numbered differently in text and refs, CityGaussian misnumbered, Mega-NeRF duplicated). These are fixable but sloppy.\n\nBottom line: the dataset and method are worth engaging with, but the load-bearing real-city claim needs much more than synthetic benchmarks. A serious referee should ask for real-world validation, proper baselines, and artifact release. I'd send it to review, not desk reject—the contribution is substantial enough—but I would not accept it in its current form.","headline":"Solid dataset and a sensible diffusion approach, but the real-city claim rests entirely on synthetic UE5 experiments and mixed metrics—worth reviewing, not accepting as is.","tokens_in":13821,"tokens_out":2229,"would_cite":false,"duration_ms":19252,"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":"Ground-level city views from aerial photos alone","keywords":["aerial-to-ground view generation","multi-view diffusion","3D urban reconstruction","3D Gaussian Splatting","novel view synthesis","city-scale dataset","privacy-preserving imaging"],"falsifier":"Take a real urban area with both drone/satellite aerial imagery and a modest ground-truth set of street-level photos, run the AerialGo pipeline to generate ground views from the aerial data, and compare the generated views and the resulting 3DGS reconstruction against the real ground-truth images. A substantial drop in PSNR/SSIM/LPIPS relative to the synthetic results would show that the method's real-city promise is not met.","tokens_in":12729,"feed_emoji":"🏙️","tokens_out":5778,"duration_ms":44648,"temperature":0.7,"pith_summary":"City-scale 3D reconstruction usually depends on collections of ground-level images, which are costly to gather and can expose license plates, faces, and other personal data. AerialGo proposes a multi-view diffusion model that generates realistic street-level (walking-through) views conditioned only on aerial imagery and a coarse point cloud derived from it. The paper claims these generated ground views, when added to standard reconstruction pipelines such as 3D Gaussian Splatting, significantly improve ground-level rendering quality. If true, detailed urban models for planning, navigation, and AR/VR could be built from aerial captures alone, sidestepping ground-level data collection and its privacy burden. The accompanying AerialGo dataset, 3.45 million paired aerial and ground images over 134 km² of synthetic cities, is released to support training of such generative models.","feed_headline":"Ground-level city views from aerial photos alone","feed_subtitle":"A diffusion model turns aerial shots into street views that improve 3D city reconstruction without ground-level data collection.","key_machinery":"The load-bearing mechanism is a multi-view latent diffusion model initialized from ImageDream and adapted to the aerial-to-ground setting. The denoising UNet processes a stack of $N=3$ encoded aerial reference latents concatenated with the noisy ground-view latent, using 3D self-attention (1D attention across views plus 2D attention within each image) so that multi-view aerial context is shared coherently. Structural conditioning comes from rendering the reconstructed 3D point cloud from the target ground camera, encoding that render with a CLIP encoder, and injecting the embeddings through cross-attention and the residual blocks, together with camera embeddings; classifier-free guidance is applied with a scale of 5.0. This design is what lets the model bridge the large angular gap between oblique aerial cameras and near-horizontal pedestrian viewpoints, and the same point-cloud conditioning is what lets the generated views stay anchored to real geometry when used as reconstruction priors.","core_discovery":"The paper's central claim is that a diffusion model can synthesize photorealistic ground-level views from aerial images, and that these synthetic views act as effective priors for 3D urban reconstruction. Trained on synthetic Unreal Engine 5 data from the new AerialGo dataset and from MatrixCity, the model takes three nearby aerial reference images and a point-cloud rendering of the target ground camera, and denoises a ground-view latent through a UNet with 3D self-attention, producing an image aligned with the requested viewpoint. When these generated images are fed into 3DGS, CityGaussian, or GigaGS along with a perceptual loss and a skybox model, the paper reports markedly better ground-level PSNR, SSIM, and LPIPS than aerial-only reconstruction, and better cross-view synthesis than prior diffusion-based novel-view methods.","pith_inferences":["The synthetic-only training means the strongest untested claim is transfer to real cities; a natural testbed is a real drone or satellite dataset with a small set of held-out street-level images for evaluation.","Because the conditioning signal is a point cloud rendered from the ground camera, the framework's real-world ceiling is set by the quality of aerial SfM/MVS; noisy or incomplete real point clouds would weaken the structural prior that the ablation shows is essential.","The same architecture could plausibly be extended from image generation to video (walking-through sequences) by replacing the 3D self-attention with temporal attention, which would move the method closer to street-level fly-through rendering.","The privacy argument extends beyond data collection: generated ground views contain no real personal data, so models trained on them could be shared or deployed in regulated settings without the anonymization overhead of real street imagery."],"forward_implications":["City-scale 3D reconstruction can be driven by aerial imagery alone; ground-level priors are produced synthetically, removing the need for dedicated ground capture campaigns.","The privacy concerns associated with street-level imagery (faces, license plates) are bypassed because no real ground images are collected or stored.","Reconstruction methods built on 3D Gaussian Splatting, NeRF, or MVS can be upgraded with a plug-in perceptual loss and skybox to exploit the generated priors, improving ground-level realism without architectural changes to the core representation.","The AerialGo dataset, with its paired aerial/ground views, depth, segmentation, and weather/lighting controls, gives the research community a training resource for cross-view generation and city-scale generative modeling that previously did not exist at this scale."],"supporting_citations":[{"why":"Frames the problem of large-scale city reconstruction from images and the reliance on abundant ground-level photographic data that AerialGo aims to replace.","marker":"[1]"},{"why":"A second city-scale dataset used with AerialGo for training and evaluation, providing the cross-dataset evidence for the method's claims.","marker":"[21]"},{"why":"The multi-view diffusion model whose pretrained weights and architecture AerialGo adapts for aerial-to-ground generation.","marker":"[44]"},{"why":"The main reconstruction backbone whose ground-level quality is shown to improve when generated priors are added.","marker":"[19]"},{"why":"A diffusion novel-view method that also conditions on point clouds; it serves as both inspiration and a comparison baseline that struggles with large viewpoint disparity.","marker":"[54]"},{"why":"Established the diffusion-prior paradigm for 3D (score distillation sampling), the general approach AerialGo transfers to city-scale reconstruction.","marker":"[33]"},{"why":"Prior satellite-to-ground city generation work that AerialGo extends to fully posed aerial-to-ground reconstruction with structural conditioning.","marker":"[24]"}],"fun_headline_variants":["Privacy-safe street views generated from aerial shots","Diffusion model synthesizes ground views from aerial images","Aerial to street: privacy-preserving city walkthroughs","Generate city views without ground-level data collection","AerialGo: walking through cities from aerial perspectives"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a diffusion model trained only on Unreal Engine 5 synthetic cities will transfer to real aerial imagery of real cities, where the reconstructed point clouds are noisier and the visual appearance differs.","fun_headline_variants_meta":{"raw":{"variants":["Privacy-safe street views generated from aerial shots","Diffusion model synthesizes ground views from aerial images","Aerial to street: privacy-preserving city walkthroughs","Generate city views without ground-level data collection","AerialGo: walking through cities from aerial perspectives"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000335,"raw_usage":{"total_tokens":1823,"prompt_tokens":878,"completion_tokens":945,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":881}},"tokens_in":494,"tokens_out":945,"duration_ms":7072,"temperature":1.0,"reasoning_tokens":881,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:07:03.188444+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real urban area with both drone/satellite aerial imagery and a modest ground-truth set of street-level photos, run the AerialGo pipeline to generate ground views from the aerial data, and compare the generated views and the resulting 3DGS reconstruction against the real ground-truth images. A substantial drop in PSNR/SSIM/LPIPS relative to the synthetic results would show that the method's real-city promise is not met.","supporting_citations":[{"cited_title":"Building rome in a day","cited_arxiv_id":null,"evidence_quote":"Frames the problem of large-scale city reconstruction from images and the reliance on abundant ground-level photographic data that AerialGo aims to replace."},{"cited_title":"Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond","cited_arxiv_id":null,"evidence_quote":"A second city-scale dataset used with AerialGo for training and evaluation, providing the cross-dataset evidence for the method's claims."},{"cited_title":"InfiniCity: Infinite-scale city synthesis","cited_arxiv_id":null,"evidence_quote":"Prior satellite-to-ground city generation work that AerialGo extends to fully posed aerial-to-ground reconstruction with structural conditioning."}],"review_version":1}