{"id":"37647656-aac4-4a5f-a2b2-8ad89724cccd","arxiv_id":"2505.23434","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"UrbanCraft uses hierarchical semantic-geometric priors to condition diffusion-based score distillation, enabling extrapolated view synthesis for urban 3D Gaussian Splatting scenes.","lead":"UrbanCraft generates new views of urban scenes from cameras looking far left, right, or down, beyond the training trajectory. It uses occupancy grids, depth, semantic maps, and object orientations to guide a diffusion model that repairs unseen areas in 3D reconstructions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on FID/KID measured against the very training images the method is optimized to match; the gap over VEGS is tiny and unreported with error bars, so the central claim is not supported.","rationale":"The reader's formal weakest_assumption points at occupancy-grid and bounding-box sensitivity, which is a genuine limitation but is not the most load-bearing issue for the central claim. The decisive problem is that the quantitative evidence for superiority is measured with a circular metric: FID/KID between EVS renders and training images, when the method's optimization objective explicitly enforces distributional consistency with those same images. The margins over VEGS are small enough to be noise, no error bars are given, and on ground-truth-view metrics the methods are on par. Consequently, the \"first effective framework\" and state-of-the-art assertions are unsupported regardless of whether the occupancy-grid assumption holds. The reader's rationale does mention evaluation quality, so there is partial agreement, but the weakest_assumption field should have been the evaluation circularity rather than occupancy sensitivity. The verdict of REJECT is appropriate because the central claim is not established, though the underlying method may still be worth investigating under a non-circular benchmark.","tokens_in":19708,"tokens_out":5243,"duration_ms":49530,"concrete_test":"Re-run UrbanCraft and VEGS on a synthetic urban dataset (e.g., CARLA) with ground-truth images rendered from the same extrapolated camera poses used for EVS; compute FID/KID between each method's EVS renders and those ground-truth images, with bootstrap 95% CIs, and verify that UrbanCraft2D was never trained on that dataset. If UrbanCraft does not beat VEGS by more than the CI width, the claimed SOTA is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (\"first effective framework to synthesize large- and text-ambiguous extrapolated camera views with consistent distribution with existing camera views\", Introduction) is supported mainly by Table 1. That table's EVS metrics compute FID/KID between the method's extrapolated renders and the training images (Sec. 4.1, 4.3). This is not an extrinsic measure of extrapolation quality: the method is explicitly optimized to make rendered views distribution-consistent with the observable/training views (L_reconst on training cameras plus HSG-VSD \"forcing the distribution to be consistent with the observable scene\"), so matching the training-image FID is close to optimizing the metric itself. It also cannot penalize hallucinated content that happens to look like the dataset. The reported margin over VEGS is tiny (FID 278.75 vs 279.27; KID 0.8906 vs 0.9245) with no error bars or significance test, so it is within plausible bootstrap noise. On conventional test cameras with ground truth, UrbanCraft is essentially on par with VEGS (PSNR 24.15 vs 24.00; LPIPS 0.219 vs 0.216), so the only claimed advantage is on the circular EVS metric. The paper's own Supp. 8.2 admits blurring from incomplete occupancy grids and Gaussian sparks in large unseen views, further limiting the scope of the headline claim. Thus the evidence does not substantiate the state-of-the-art claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UrbanCraft, a framework for extrapolated view synthesis (EVS) in urban scenes. It introduces a pretrained 2D diffusion model (UrbanCraft2D) conditioned on hierarchical semantic-geometric priors (occupancy grids, semantic/depth maps, and instance-level rotation maps), and distills this model into a 3D Gaussian representation via a proposed HSG-VSD loss. The authors claim state-of-the-art EVS performance, reporting quantitative results on KITTI-360 and qualitative results on KITTI-360 and NuScenes.","tokens_in":20024,"tokens_out":4517,"duration_ms":39494,"significance":"The work addresses a practically important problem: rendering urban scenes from viewpoints far outside the training camera distribution. The method is well motivated, combining scene-level and instance-level control signals in a single diffusion prior, and the qualitative results show plausible repair of unseen regions, including improved vehicle geometry and color consistency. The paper also provides extensive supplementary material, including failure cases and public-resource acknowledgments. However, the central quantitative claim of state-of-the-art performance is not well supported: the main EVS metrics (FID/KID) are computed against the very training images used to fine-tune UrbanCraft2D (Supp. 6.1), the margins over VEGS are tiny and lack error bars, and on conventional test-view metrics the method is on par with VEGS. The paper's own limitations (Supp. 8.2) admit blurring from incomplete occupancy grids and Gaussian sparks in large unseen views, further tempering the headline claim.","major_comments":[{"comment":"The central state-of-the-art claim rests on FID/KID computed between EVS renders and training images, but UrbanCraft2D is pretrained on the same KITTI-360 and NuScenes datasets used for evaluation (Supp. 6.1). This makes the EVS metric close to circular: the diffusion prior has seen the test scenes, and the HSG-VSD loss explicitly forces distribution consistency with the observable/training views, so matching FID/KID is essentially optimizing the metric itself. The reported gap over VEGS is tiny (FID 278.75 vs 279.27; KID 0.8906 vs 0.9245) and is reported without error bars or significance testing. On conventional test views with ground truth (Table 1), UrbanCraft is essentially on par with VEGS (PSNR 24.15 vs 24.00; LPIPS 0.219 vs 0.216). The evidence therefore does not substantiate the state-of-the-art claim. Please provide held-out-scene or cross-dataset evaluation, error bars over multiple runs, and/or metrics that do not rely on the training distribution.","section":"Sec. 4.3, Table 1; Supp. 6.1"},{"comment":"The method assumes accurate occupancy grids and 3D bounding boxes as inputs, and the authors' own failure analysis (Supp. 8.2) states that incomplete occupancy grids cause blurring artifacts and that large unseen views produce Gaussian sparks due to an optimization imbalance. These are not merely edge cases: they directly affect the claimed capability to synthesize large- and text-ambiguous extrapolated camera views. The paper should quantify how often these failures occur (e.g., by difficulty level or scene) and discuss whether the hierarchical priors actually improve robustness to imperfect annotations, or at least temper the claim accordingly.","section":"Sec. 3.1; Supp. 8.2"},{"comment":"The paper mentions the existing EVS benchmark [71] only in a supplementary discussion and does not report results on that benchmark or compare with other EVS baselines beyond VEGS and 3DGS variants. Given the paper's title and central claim, a direct comparison on a standardized benchmark with fixed train/test splits is necessary to support the state-of-the-art assertion. Without it, the evidence is limited to a single dataset (KITTI-360) under the authors' own evaluation protocol.","section":"Supp. 6.6; Table 1"}],"minor_comments":[{"comment":"The caption says 'HSD-VSD distillation'; this should be 'HSG-VSD distillation'.","section":"Table 1 caption"},{"comment":"The caption says 'sharp cone phonemes'; this appears to be a typo for 'sharp cone phenomena'.","section":"Figure 8 caption"},{"comment":"The expression 'epsilon_p(x_t, t, y, psi(SG(T)) - epsilon_phi(...))' has an ambiguous parenthesis structure; a closing parenthesis is needed before the minus sign for clarity.","section":"Eq. (1)"},{"comment":"The phrase 'extent it to a more comprehensive setting' appears to be a typo for 'extend it to a more comprehensive setting'.","section":"Sec. 4.1"},{"comment":"The term 'sem-geometric' is used in the abstract but not defined until Section 3.1; please define it at first use.","section":"Abstract / Sec. 3.1"}],"recommendation":"major_revision","confidential_remarks":"The evaluation circularity is the main concern; any revision must address it with held-out experiments. The tiny FID gap and the admitted failure cases make the 'first effective framework' claim difficult to defend. I would be open to a revised version that restricts the claims to the qualitative improvements and provides a non-circular evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"UrbanCraft is a genuinely new combination: it conditions a ControlNet-based diffusion prior on occupancy grids, semantic maps, depth, and instance rotation maps, and then uses that as a VSD-style distillation target for 3DGS urban scenes. That is a sensible way to attack the text-ambiguity problem in extrapolated view synthesis, and the qualitative figures show real differences from VEGS, especially on vehicle geometry and large-angle views. The ablations (instance-level control, G-SDS) are in the right direction, and the supplement is unusually honest about failure modes: incomplete occupancy grids cause blur, large unseen views cause Gaussian sparks.\n\nThe problem is the evaluation supports only weak claims. UrbanCraft2D is pretrained on KITTI-360 and NuScenes, the same datasets used for evaluation. The main metric, FID/KID, is computed against the training images of those scenes, and the model is explicitly optimized to make its distribution consistent with those same training images. So the 0.52 FID gap over VEGS is close to optimizing the metric itself. No error bars or significance tests are reported, and on conventional test views with ground truth, UrbanCraft is essentially tied with VEGS (PSNR 24.15 vs 24.00; LPIPS slightly worse). The qualitative results are suggestive, but they are selected frames, and the paper does not compare against a version of VEGS that also gets the same occupancy/box priors. So the headline claim of \"first effective framework\" and state-of-the-art performance is not substantiated.\n\nI would not call this a desk-reject. The framework is plausible, the conditioning representation is well thought out, and the problem is important. But the paper needs either (a) a held-out evaluation protocol, e.g. fine-tuning UrbanCraft2D on one city and testing on a disjoint city, with FID computed against a held-out set or against ground-truth extrapolated views if available, or (b) a clear statement that the contribution is a method plus qualitative improvements, with the quantitative SOTA claim removed. Also, error bars and a full ablation table with numbers would help.\n\nFor a reader: the paper is worth discussing in a reading group on generative NVS, mostly as an example of how to design scene-level conditioning, and as a cautionary example of dataset-circular evaluation. I wouldn't cite it in my own work yet. But I would send it to peer review—the idea deserves referee attention, and a serious referee could push the authors toward a fair evaluation. My own verdict would be major revision, not reject, if the evaluation gets fixed.","headline":"A clever, plausible conditioning scheme for urban extrapolated view synthesis, but the SOTA claim rests on a dataset-circular FID metric with a tiny margin, so the evidence supports only a qualitative advance.","tokens_in":20575,"tokens_out":1736,"would_cite":false,"duration_ms":17209,"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":"UrbanCraft claims the first framework that synthesizes large and text-ambiguous extrapolated urban views with distribution consistent with the observed scene, guided by hierarchical semantic-geometric priors.","keywords":["extrapolated view synthesis","urban scene reconstruction","3D Gaussian Splatting","diffusion model","score distillation sampling","occupancy grid","semantic-geometric priors","ControlNet"],"falsifier":"Render an extrapolated view in a scene whose occupancy grid has been deliberately truncated, for example cutting the perception range in half, while keeping the true scene otherwise unchanged; if blurring artifacts appear exactly in the truncated regions and persist after HSG-VSD optimization, then the claim that hierarchical sem-geometric priors enable consistent large extrapolation fails for exactly the regime the paper targets.","tokens_in":19499,"feed_emoji":"🏙️","tokens_out":7239,"duration_ms":71614,"temperature":0.7,"pith_summary":"The paper takes on Extrapolated View Synthesis (EVS) for urban scenes reconstructed from forward-facing vehicle cameras: synthesizing photorealistic images from viewpoints far outside the training camera trajectory, such as looking left, right, or downward. It claims to be the first effective framework for large and text-ambiguous extrapolated views, where previous text-only diffusion guidance fails because a text prompt cannot specify local layout, occlusion, or orientation. UrbanCraft's answer is to condition a pretrained text-to-image diffusion model on hierarchical semantic-geometric priors, a scene-level occupancy grid rendered at the extrapolated pose plus instance-level 3D bounding-box rotation maps, and to distill the conditioned model into a 3D Gaussian Splatting scene through a new Hierarchical Sem-Geometric-Guided Variational Score Distillation (HSG-VSD) loss. If correct, this makes urban reconstructions usable from viewpoints no training image ever saw, and turns coarse semantic-geometric maps into enough control to keep generated content consistent with the observed scene.","feed_headline":"Street scenes render cleanly from far-off camera angles","feed_subtitle":"Occupancy grids and 3D boxes steer diffusion so extrapolated views match the observed scene.","key_machinery":"The load-bearing object is the Hierarchical Sem-Geometric (HSG) representation and its use in score distillation. HSG is a control signal assembled by rendering a scene-level occupancy grid, giving semantic and depth maps, and an instance-level projected rotation map, the flattened 3x3 rotation matrix of each object's 3D bounding box, at the extrapolated camera pose, yielding a 13-channel condition. This signal is fed through ControlNet into a Stable Diffusion v2.1 model finetuned on urban datasets as UrbanCraft2D; a fixed base prompt is used so text does not dominate. HSG-VSD then replaces the text-only score distillation gradient with a gradient comparing the ControlNet-conditioned denoiser against a LoRA-adapted denoiser, both conditioned on the rendered sem-geometric signal, so that each optimization step pulls the rendered image toward the distribution of the observable scene. A supplementary geometry score distillation loss on rendered depth and normals regularizes Gaussian covariance shapes to suppress surface cavities and floating artifacts.","core_discovery":"UrbanCraft claims that the missing ingredient for urban extrapolated view synthesis is not a bigger or better diffusion model but hierarchical semantic-geometric control. The paper shows that scene-level occupancy grids, providing coarse semantic and depth layout, together with instance-level projected rotation maps from 3D bounding boxes, providing object orientation and spatial relationships, form a 13-channel control signal that can be injected into a finetuned Stable Diffusion model via ControlNet. This conditioned model, named UrbanCraft2D, is then distilled into a static-plus-dynamic 3D Gaussian Splatting scene using HSG-VSD, a variational score distillation loss in which ControlNet features condition both the pretrained denoiser and a LoRA-adapted denoiser at each extrapolated camera pose. The paper reports state-of-the-art FID and KID on extrapolated views while keeping on-par quality on conventional test cameras, and presents this as the first framework that repairs large, text-ambiguous extrapolated views without collapsing vehicle geometry or losing color and depth consistency with the observed scene.","pith_inferences":["The single biggest lever on final quality may be the completeness of the occupancy grid rather than the diffusion model itself: the paper's own failure analysis ties blurring to incomplete grids, which suggests a testable recipe of corrupting the grid, such as truncating its perception range, and measuring how FID degrades.","The HSG recipe is domain-generic: any setting where cheap geometric primitives, such as maps, boxes, or depth, can be rendered from a novel pose could use the same ControlNet-plus-variational-score-distillation conditioning, so the approach should transfer to indoor navigation, aerial or satellite view synthesis, or other structured scenes, not just street driving.","Because the paper evaluates distribution consistency with FID and KID relative to training images, those numbers do not directly measure geometric correctness at extrapolated views; a stronger public test would hold out ground-truth images from a parallel vehicle pass and measure pixel-level error there.","The reported speed of about two iterations per second, together with the ablation showing the SDS-style variant suffers over-saturation and over-smoothing, suggests the variational branch is the accuracy bottleneck and that a faster SDS-style variant may be the practical path to real-time urban scene repair."],"forward_implications":["Extrapolated views that are far left, far right, or downward can be rendered without the black holes, blurring, and collapsing vehicle geometry that text-only diffusion baselines produce.","Text ambiguity stops being the main failure mode because local layout comes from occupancy grids and object poses, so a single prompt can guide many different extrapolated views consistently.","With scene-completion algorithms supplying occupancy grids and detectors providing 3D boxes, the framework can be applied to custom driving datasets that lack dense multi-view imagery.","Instance-level control is strong enough to support editing, such as changing a car's color while preserving its geometry and position, which the paper demonstrates as an additional capability.","The two-stage training recipe, reconstruction first and distillation repair second, plus a LoRA-based style transfer, lets a model pretrained on one urban dataset be adapted to another."],"supporting_citations":[{"why":"Supplies the text-only diffusion EVS baseline and the extrapolated-camera evaluation protocol and FID/KID metrics that UrbanCraft must beat.","marker":"[26]"},{"why":"Provides the ControlNet mechanism that injects the 13-channel hierarchical sem-geometric control signal into the diffusion model.","marker":"[30]"},{"why":"Defines the 3D Gaussian Splatting representation that UrbanCraft initializes and optimizes with reconstruction and distillation losses.","marker":"[38]"},{"why":"Supplies the Stable Diffusion v2.1 base text-to-image model that is finetuned into UrbanCraft2D.","marker":"[54]"},{"why":"Introduces score distillation sampling, the optimization method from which HSG-VSD and the geometry SDS loss are derived.","marker":"[59]"},{"why":"Introduces variational score distillation, the framework that HSG-VSD extends with ControlNet conditioning on sem-geometric priors.","marker":"[61]"},{"why":"Provides the KITTI-360 forward-facing urban dataset used for pretraining UrbanCraft2D and for the main EVS benchmarks.","marker":"[23]"},{"why":"Provides the nuScenes multi-view urban dataset used for pretraining UrbanCraft2D and for cross-dataset validation.","marker":"[31]"},{"why":"Supplies the Occ3D occupancy grid maps for nuScenes that are rendered into scene-level control signals.","marker":"[28]"},{"why":"Supplies the SSCBench occupancy grid maps for KITTI-360 that are rendered into scene-level control signals.","marker":"[29]"}],"fun_headline_variants":["UrbanCraft breaks view limits with geometric priors","Extrapolated street views tamed by occupancy grids and 3D boxes","Sem-geometric priors power UrbanCraft's far-view synthesis","Steering diffusion with 3D boxes for urban view extrapolation","UrbanCraft: extrapolating city views via hierarchical priors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that accurate occupancy grid maps and 3D bounding-box annotations for the scene are available and can be rendered correctly at the extrapolated camera pose; if those maps are incomplete or wrong, the control signal is wrong and the whole pipeline degrades, which the paper's own limitation section confirms by tying blurring artifacts to incomplete occupancy grids.","fun_headline_variants_meta":{"raw":{"variants":["UrbanCraft breaks view limits with geometric priors","Extrapolated street views tamed by occupancy grids and 3D boxes","Sem-geometric priors power UrbanCraft's far-view synthesis","Steering diffusion with 3D boxes for urban view extrapolation","UrbanCraft: extrapolating city views via hierarchical priors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000291,"raw_usage":{"total_tokens":1734,"prompt_tokens":1012,"completion_tokens":722,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":632}},"tokens_in":628,"tokens_out":722,"duration_ms":7174,"temperature":1.0,"reasoning_tokens":632,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:45:14.527406+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render an extrapolated view in a scene whose occupancy grid has been deliberately truncated, for example cutting the perception range in half, while keeping the true scene otherwise unchanged; if blurring artifacts appear exactly in the truncated regions and persist after HSG-VSD optimization, then the claim that hierarchical sem-geometric priors enable consistent large extrapolation fails for exactly the regime the paper targets.","supporting_citations":[{"cited_title":"Vegs: View extrapolation of urban scenes in 3d gaussian splatting using learned priors","cited_arxiv_id":null,"evidence_quote":"Supplies the text-only diffusion EVS baseline and the extrapolated-camera evaluation protocol and FID/KID metrics that UrbanCraft must beat."},{"cited_title":"High-resolution image synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Supplies the Stable Diffusion v2.1 base text-to-image model that is finetuned into UrbanCraft2D."},{"cited_title":"Dreamfusion: Text-to-3d using 2d diffusion","cited_arxiv_id":null,"evidence_quote":"Introduces score distillation sampling, the optimization method from which HSG-VSD and the geometry SDS loss are derived."},{"cited_title":"Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distilla- tion","cited_arxiv_id":null,"evidence_quote":"Introduces variational score distillation, the framework that HSG-VSD extends with ControlNet conditioning on sem-geometric priors."},{"cited_title":"Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d","cited_arxiv_id":null,"evidence_quote":"Provides the KITTI-360 forward-facing urban dataset used for pretraining UrbanCraft2D and for the main EVS benchmarks."},{"cited_title":"nuscenes: A multi- modal dataset for autonomous driving","cited_arxiv_id":null,"evidence_quote":"Provides the nuScenes multi-view urban dataset used for pretraining UrbanCraft2D and for cross-dataset validation."},{"cited_title":"Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving","cited_arxiv_id":null,"evidence_quote":"Supplies the Occ3D occupancy grid maps for nuScenes that are rendered into scene-level control signals."},{"cited_title":"Ss- cbench: A large-scale 3d semantic scene completion bench- mark for autonomous driving","cited_arxiv_id":null,"evidence_quote":"Supplies the SSCBench occupancy grid maps for KITTI-360 that are rendered into scene-level control signals."}],"review_version":1}