Pith. sign in

REVIEW 5 major objections 6 minor 3 cited by

Sat2RealCity turns satellite imagery into geometry-aware, style-consistent 3D urban scenes at building level.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 22:10 UTC pith:TLZKMUEC

load-bearing objection Serious engineering effort, but the city-scale geometry numbers are circular: OSM footprints/heights are both conditioning input and evaluation ground truth. the 5 major comments →

arxiv 2511.11470 v3 pith:TLZKMUEC submitted 2025-11-14 cs.CV

Sat2RealCity: Geometry-Aware and Appearance-Controllable 3D Urban Generation from Satellite Imagery

Xinliang Wang , Yijie Kang , Zhenyu Wu , Yifeng Shi This is my paper
classification cs.CV
keywords 3D urban generationsatellite imageryOpenStreetMapbuilding-entity generationsparse structure latentappearance controlmultimodal LLM3D building dataset
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that city-scale 3D generation can be decomposed into per-building generation, reusing pretrained object-level 3D generators, with geometry grounded in OSM footprints and heights and appearance controlled by multimodal-language-model-written style descriptions. If this holds, digital twins and simulation environments can be built from public satellite and map data alone, without expensive 3D city assets. The central evidence is that generation initialized from an OSM-derived latent prior matches footprints in position and structure, while a shared frontal appearance image keeps regional facades coherent.

Core claim

The core claim is that injecting OSM-derived structural priors into a sparse-structure latent—by interpolating the encoded extrusion with Gaussian noise—makes a fine-tuned building generator produce 3D shapes that align with real-world footprints in both geometry and geospatial pose, and that dual cross-attention on top-view and frontal-view images separates structure from style so a region's buildings share a coherent appearance while retaining individual geometries. Experiments report lower Chamfer distance and higher F-score against OSM-based ground truth, plus higher CLIP scores over city-generation baselines.

What carries the argument

The central mechanism is the fused geometric prior: OSM footprints extruded to volumetric proxies, encoded into a Sparse Structure (SS) latent space (where each latent marks activated voxels on a surface), normalized, and mixed with Gaussian noise via cosine interpolation (Eq. 2). This fused latent initializes the structure-flow transformer, replacing pure noise. Parallel cross-attention pathways condition on the satellite top view (structure) and an MLLM-generated frontal image (appearance), and the model learns to weight roof cues from the top pathway and facade cues from the appearance pathway.

Load-bearing premise

The load-bearing premise is that OSM footprint-and-height data are an accurate measure of real building geometry; if that data is coarse, outdated, or wrong, the reported geometric scores largely reflect the model copying its own conditioning rather than measuring true accuracy.

What would settle it

Evaluate the method on a region with independent LiDAR or photogrammetric ground truth, where OSM footprints are known to be outdated or deliberately corrupted; if the generated geometry tracks the corrupted OSM data and fails against the independent ground truth, the geometric alignment claim reduces to prior replication.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The method can produce 3D cities from only satellite imagery and OSM data, removing the need for licensed or manually built 3D city datasets.
  • Geospatial pose alignment is learned into the generation process, so post-hoc rotational registration of generated buildings becomes unnecessary.
  • A single shared frontal appearance image can control style for an entire neighborhood without sacrificing per-building footprint accuracy.
  • The framework can accept real street-view imagery as appearance input, enabling appearance grounding in actual urban photographs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because OSM footprints and heights are used both as generation conditioning and as evaluation ground truth, the reported geometric gains may partly reflect the model copying its own conditioning signal rather than learning true building geometry; an independent ground-truth comparison would separate the two.
  • The appearance maps are generated from text descriptions, so swapping the text-to-image step for real facade photos could let the system reflect actual street appearance without retraining the 3D model.
  • The building-entity decomposition suggests a divide-and-conquer strategy for unbounded city generation: each building can be generated in parallel and assembled at OSM coordinates, scaling to arbitrarily large regions.
  • Other height or footprint sources (e.g., LiDAR-derived) could replace OSM data in the prior, potentially improving accuracy where OSM is sparse or outdated.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes Sat2RealCity, a building-entity-based 3D urban generation framework that takes satellite imagery and OSM data as input. The method injects OSM footprints and heights as structural priors into a TRELLIS-style latent diffusion model, adds dual cross-attention pathways for top-view and frontal-view appearance conditioning, and uses an MLLM pipeline to generate regional appearance descriptions. The authors construct a 11,579-building AIGC dataset, fine-tune the model, and evaluate on 111 real-world regions totaling 10 km². They report that Sat2RealCity outperforms object-level 3D generation baselines and city-level generators in Chamfer Distance, F-score, CLIP score, and a regional consistency score, and they claim strong geographic alignment, regional stylistic consistency, and plausible urban asset synthesis.

Significance. If the claims are valid, the paper offers a practical route to city-scale 3D content from public geodata and satellite imagery without large 3D city assets. The entity-level decomposition and the reuse of a pretrained object-level 3D generator are sensible design choices, and the planned release of the BuildVerse3D dataset and code is a positive contribution. However, the central quantitative evidence is currently not probative because the geometry evaluation uses OSM-derived footprints and heights as ground truth while the method injects the same OSM data as its conditioning prior. The ablation on 500 CityCraft assets partly addresses this, but it is building-level and may still leak the test footprint through the structural prior. The appearance and style metrics are also weak: CLIP-based scores and within-generated-set pairwise similarity do not establish faithfulness to real-world appearance. The framework is promising, but the experiments need substantial rework before the central claims are supported.

major comments (5)
  1. [§3.3 (Eq. 2) vs §4.2 (Table 1)] The geometry evaluation is structurally circular. §4.2 defines ground-truth geometry as OSM building footprints with height information, while §3.3 (Eq. 2) injects the same OSM footprints/heights as the structural prior Z''_O, with λ=0.5 and LOD 1 used at inference (Sec. 4.1). Thus CD and F-score in Table 1 largely measure how faithfully the generator copies its conditioning latent. Baselines do not receive these OSM priors, so the comparison is asymmetric and does not establish real-world geometric fidelity. A direct extrude-the-prior baseline should be added, and the evaluation should be repeated against independent geometry such as LiDAR or photogrammetric meshes, or at least against OSM heights for held-out regions with a clear statement of what is and is not being measured.
  2. [§4.4 / Table 2] The CityCraft ablation is not fully free of the leakage problem. Sec. 3.3 says the OSM prior during training is 'obtained from a downsampled version of a 3D asset voxel'; if the test assets' own footprints or heights are used to construct the structural prior, then the prior encodes the answer for CD/F even though the ground-truth mesh is a full 3D model. The paper should state explicitly whether the test footprints/heights are inputs at test time, and should include an ablation where the prior is derived only from the satellite image (e.g., a predicted height map) or where the prior is intentionally corrupted to quantify sensitivity. Without this, the claimed superiority of Sat2RealCity-T over baselines in Table 2 may reflect prior injection rather than geometric reconstruction skill.
  3. [§4.5 / Eq. (5)] The Regional Consistency Score S_regional = IoU_top × CLIP_pairwise is partly circular and partly self-referential. The IoU_top term measures alignment with OSM footprints, which are the same footprints used as the structural prior in §3.3. The CLIP_pairwise term computes similarity among generated buildings in the same region, but does not compare against the real region's appearance; a method that ignores the satellite images and applies a single generic facade to all buildings could score high on CLIP_pairwise as long as IoU_top is satisfied. A reference-based style metric, such as CLIP or FID against Google Earth renderings of the same area, and/or a human perceptual study, is needed to support the claim of regional stylistic consistency.
  4. [§4.6 / Table 4] The comparison with city generation methods is not sufficiently controlled. CityDreamer outputs only rendered RGB frames without explicit 3D geometry, so only CLIP is reported; Seed3D does not accept satellite inputs and is instead fed Google Earth Studio oblique images, making the input conditions fundamentally different. The text claims 'superior geometric structure' for Sat2RealCity from Figure 7, but no geometry metric is reported for this comparison. This experiment supports only a narrow claim about CLIP-based appearance similarity, and the discussion overstates what it demonstrates.
  5. [§3.3, abstract claims] The claim that 'every generated building precisely matches its satellite footprint in both geometric structure and geospatial pose' is stronger than what the metrics measure. CD and F-score are whole-region point-cloud distances to an OSM extrusion; they do not verify per-building footprint polygon accuracy, height accuracy against independent measurements, or orientation beyond what is implicitly captured by the footprint prior. The paper should report a per-building metric such as footprint IoU and height error against an independent source, and should temper the abstract/claims accordingly until such evidence is provided.
minor comments (6)
  1. [Abstract / front matter] The manuscript contains two different abstracts: the one in the arXiv metadata and the one at the beginning of the full text differ substantially in wording and emphasis. This looks like a version-control error and should be corrected.
  2. [§3.2 and §3.5] Commercial components such as 'Google Imagen 4' in §3.2 and the Overture/Mapbox data sources in §4.1 are mentioned without proper references or version numbers. Please add citations and access dates, and clarify the license/terms for the released dataset if these tools are used.
  3. [§4.1 / Eq. (2)] λ is a free hyperparameter set to 0.5, but the paper does not report sensitivity to it. Given that the structural prior is the main novelty, a small sweep or at least a stability statement would strengthen the paper.
  4. [§4.3 / Table 1] TRELLIS-MV-M and TRELLIS-MV-S are described only as 'multi-diffusion-based' and 'stochastic sampling'. Please give a one-sentence definition of each, or cite the source, so readers can understand the difference.
  5. [§4.4 / Table 2] The text says the held-out set consists of '500 high-quality building assets manually modeled in the CityCraft [5]'. CityCraft is a generation method, not a dataset; please clarify what CityCraft means here and describe the manual modeling process or provide a citation to the actual dataset.
  6. [General] No error bars or statistical significance tests are reported for any metric. Given the small number of regions (a subset of 20 in §4.5) and the large reported margins, standard deviations or confidence intervals would help assess robustness.

Circularity Check

2 steps flagged

Regional geometry evaluation is circular: OSM footprints/heights are both the conditioning input (Eq. 2) and the ground-truth geometry (Sec. 4.2), so CD/F can be satisfied by copying the conditioning latent; the style-consistency metric also rewards the shared frontal input.

specific steps
  1. self definitional [Sec. 3.3 (Eq. 2), Sec. 4.1, Sec. 4.2 / Table 1]
    "We first extrude OSM building footprints into coarse volumetric proxies ... Z''_O = cos(λπ/2) Z'_O + sin(λπ/2) ε ... The hyperparameter λ ... was set to 0.5, and only LOD 1 priors were used. Ground truth data are derived from OSM building footprints with height information for geometry."

    The same OSM footprints plus heights that are encoded into the SS-latent initialization (Eq. 2) are used as the geometric ground truth for CD/F-score in Table 1. With λ=0.5 and LOD 1 at inference, the generator is initialized from a strong copy of the evaluation target; low CD/high F may mostly measure reproduction of the conditioning latent. Baselines do not receive this OSM prior, so the comparison is asymmetric and does not establish real-world reconstruction accuracy.

  2. self definitional [Sec. 4.5 / Eq. 5, Table 3]
    "For each region, a frontal image from our MLLM pipeline served as the shared style prompt, and each building was generated from its unique satellite top-view input combined with this common appearance condition. ... Appearance Consistency (CLIP pairwise): ... average CLIP (ViT-L/14) cosine similarity is computed over all unique N(N−1)/2 pairs. Sregional = IoU_top × CLIP_pairwise."

    The regional consistency score rewards the exact mechanism the method installs: all buildings in a region are conditioned on the same frontal appearance image, so high pairwise CLIP similarity is partly a direct consequence of the shared input. The IoU_top factor compares against OSM footprints, which are also injected as the structural prior, further re-scoring the conditioning signal rather than providing an independent estimate of regional style quality.

full rationale

The derivation chain itself is not circular: Sat2RealCity fine-tunes pretrained TRELLIS with an OSM-derived SS-latent initialization and dual cross-attention conditions, and the building-level geometric ablation on 500 manually modeled CityCraft assets (Sec. 4.4) is an independent, non-circular validation. However, the paper's strongest regional claim—geometric and geospatial fidelity at city scale—is supported mainly by Table 1, where the evaluation ground truth is the OSM footprint+height extrusion that is simultaneously used as the generation prior. The regional style metric in Sec. 4.5 similarly measures pairwise similarity across buildings that are deliberately given the same frontal appearance condition. These evaluations make part of the headline result true by construction rather than by independent measurement. There is no load-bearing self-citation or author-imported uniqueness theorem; the circularity is in the evaluation design. Score 7 reflects one central prediction class reducing to its conditioning input, with a separate non-circular building-level ablation preventing a higher score.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced; the main inventions are a pipeline and a dataset, captured above under free parameters and domain assumptions.

free parameters (4)
  • lambda (structural prior interpolation strength) = 0.5
    Chosen hyperparameter in Eq. 2 balancing OSM prior latent Z'_O against Gaussian noise. Inference uses 0.5 and only LOD1 priors; directly controls how strongly geometry is forced to match OSM.
  • F-score distance threshold = not reported
    F-score depends on a distance threshold stated in Sec. 4.2 but never given; threshold choice substantially changes the score and prevents independent comparison.
  • HDBSCAN clustering hyperparameters = not reported
    Building clustering for shared style in Sec. 3.5 depends on HDBSCAN parameters (min_cluster_size, etc.); not specified, affecting style grouping and downstream consistency.
  • Training LOD sampling schedule = random LOD0/LOD1 in training; LOD1 at inference
    Design choice in Sec. 3.3 and Sec. 4.1; the pairing of LOD and noise intensity affects generalization and reported geometry metrics.
axioms (4)
  • domain assumption OSM footprints and Overture height attributes are accurate, current, and detailed enough to serve as both conditioning priors and geometric ground truth.
    Relied on in Sec. 3.3 for structural priors and Sec. 4.2 for GT geometry; if OSM is coarse or outdated, the reported CD/F do not measure real-world geometry.
  • domain assumption AIGC text-to-image-to-3D building models (Google Imagen + Hunyuan3D 2.5) are representative of real-world building geometry and appearance for fine-tuning.
    The BuildVerse3D dataset in Sec. 3.2 is entirely synthetic; no validation that the synthetic distribution covers real urban morphology and material diversity.
  • domain assumption The MLLM (Qwen3-VL) can infer true facade appearance from top-view satellite images, and FLUX-dev can convert descriptions into useful frontal reference images.
    The main appearance guidance in Sec. 3.5 is generated rather than measured; if MLLM descriptions are hallucinated, appearance 'grounding' is not established.
  • domain assumption Pretrained TRELLIS SS/SLAT latents and decoders transfer to building-entity latents and remain well-calibrated under OSM latent interpolation.
    This is the foundation of Sec. 3.3–3.4; the paper provides no analysis that injected OSM latents stay on the learned latent manifold.

pith-pipeline@v1.3.0-alltime-deepseek · 13099 in / 17474 out tokens · 151445 ms · 2026-08-03T22:10:28.124051+00:00 · methodology

0 comments
read the original abstract

3D urban generation from satellite imagery is an important task for scalable digital twins and real-world simulation environments. Existing approaches primarily rely on scene-level generation paradigms, which often require large-scale 3D city assets and struggle with controllability, geographic alignment, and realistic appearance grounding in real-world urban environments. To address these limitations, we present Sat2RealCity, a grounded urban generation framework that leverages object-level 3D generative priors for scalable city synthesis from satellite imagery. Our framework decomposes cities into geographically grounded building entities, enabling the reuse of pretrained object-level 3D generative priors while preserving real-world spatial structures. Supported by our constructed BuildVerse3D dataset, (1) we introduce an OpenStreetMap (OSM)-guided spatial grounding strategy to inject geospatial constraints into the 3D generation process; (2) we design an appearance-guided controllable generation mechanism for realistic architectural appearance and regional style consistency; and (3) we construct an MLLM-powered semantic pipeline for regional appearance understanding and semantic-aware appearance synthesis. Extensive experiments demonstrate that Sat2RealCity achieves strong geographic alignment, regional stylistic consistency, and plausible urban asset synthesis compared with existing urban generation and 3D asset generation approaches.

Figures

Figures reproduced from arXiv: 2511.11470 by Xinliang Wang, Yifeng Shi, Yijie Kang, Zhenyu Wu.

Figure 1
Figure 1. Figure 1: We present Sat2RealCity, a novel framework for gen￾erating high-fidelity 3D city models with detailed geometry and appearance from real-world satellite imagery. 49]. Such advances have opened up new opportunities for constructing digital twins, virtual cities, and large-scale simulation environments. In particular, 3D urban genera￾tion is the task of producing realistic and structurally con￾sistent 3D city… view at source ↗
Figure 2
Figure 2. Figure 2: The overview of Sat2RealCity. (a) The OSM-based Structural Priors Strategy converts OSM data into a fused geometric prior Z ′′ O. (b) The Appearance-guided Modeling Mechanism uses Z ′′ O, the top view feature ct, and a frontal appearance image feature cf to generate the 3D building. (c) The MLLM-powered Generation Pipeline provides geometric priors and the frontal-view image for modules (a) and (b). Finall… view at source ↗
Figure 3
Figure 3. Figure 3: Examples of the 3D Building Dataset. style, wall material, roof type) and a rule-based combina￾tion scheme. A depth-first search (DFS) traverses the com￾ponent tree, filtering incompatible attribute combinations (e.g., excluding glass curtain walls for industrial buildings) to ensure semantic and structural coherence. Each gener￾ated prompt is validated by a large language model for natu￾ralness and plausi… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative visualization of 3D urban generation. steps on 8 NVIDIA A100 GPUs. During training, the OSM￾based structural priors module randomly adopted LOD 0 or LOD 1. During inference, satellite imagery and OSM height data were collected from Mapbox and the Overture Maps Foundation, respectively. The clustering module uti￾lized HDBSCAN. The MLLM responsible for generating textual appearance descriptions w… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative visualization of stylistic consistency. ometry and height proportions. In contrast, our method re￾constructs sharp, dimensionally consistent structures across diverse building types. These qualitative results further sup￾port the quantitative findings, highlighting that the OSM￾based prior serves as a strong geometric anchor for single￾view 3D generation. 4.5. Ablation Study on Stylistic Consis… view at source ↗
Figure 8
Figure 8. Figure 8: Extended application with real-world street views. and Seed3D, we evaluate appearance, where our method achieves the highest CLIP score, as shown in [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. SatSplatDiff: Geometry-preserving generative refinement for high-fidelity satellite Gaussian Splatting

    cs.CV 2026-06 unverdicted novelty 6.0

    SatSplatDiff combines depth supervision and shadow-guided generative refinement with 2DGS to reduce geometric MAE by up to 18% and improve visual fidelity by 28-45% on satellite datasets while enabling 5x resolution e...

  2. SatSplatDiff: Geometry-preserving generative refinement for high-fidelity satellite Gaussian Splatting

    cs.CV 2026-06 unverdicted novelty 5.0

    SatSplatDiff combines depth supervision and shadow-guided generative refinement with prior SatSplat geometry to reduce hallucinations while raising visual quality in satellite 3D reconstruction.

  3. Sat2City v2: Native 3D City Asset Generation from a Single Satellite Image

    cs.CV 2026-06 unverdicted novelty 5.0

    Sat2City v2 adapts a pretrained native 3D latent model to generate controllable textured 3D city assets from satellite images via geometry flow fine-tuning and anchored texturing on a collected real dataset.

Reference graph

Works this paper leans on

53 extracted references · 25 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Citydreamer4d: Compositional generative model of un- bounded 4d cities.arXiv e-prints, pages arXiv–2501, 2025. 3

    2025

  2. [2]

    Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhao- hai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Jun- yang Lin. Qwen2.5-vl technical repor...

    Pith/arXiv arXiv 2025

  3. [3]

    Abo: Dataset and benchmarks for real-world 3d object un- derstanding

    Jasmine Collins, Shubham Goel, Kenan Deng, Achlesh- war Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object un- derstanding. InCVPR, pages 21126–21136, 2022. 3

    2022

  4. [4]

    Objaverse-xl: A universe of 10m+ 3d objects.arXiv preprint arXiv:2307.05663, 2023

    Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Chris- tian Laforte, Vikram V oleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl V ondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objaverse-xl: A universe of 10m+ 3d objects.arXiv preprint arXiv:2307.05663, 2023. 3

    Pith/arXiv arXiv 2023

  5. [5]

    Citycraft: A real crafter for 3d city generation.arXiv preprint arXiv:2406.04983, 2024

    Jie Deng, Wenhao Chai, Junsheng Huang, Zhonghan Zhao, Qixuan Huang, Mingyan Gao, Jianshu Guo, Shengyu Hao, Wenhao Hu, Jenq-Neng Hwang, et al. Citycraft: A real crafter for 3d city generation.arXiv preprint arXiv:2406.04983, 2024. 1, 2, 7

    Pith/arXiv arXiv 2024

  6. [6]

    Citygen: Infinite and control- lable city layout generation

    Jie Deng, Wenhao Chai, Jianshu Guo, Qixuan Huang, Junsheng Huang, Wenhao Hu, Shengyu Hao, Jenq-Neng Hwang, and Gaoang Wang. Citygen: Infinite and control- lable city layout generation. InCVPR, pages 1995–2005,

    1995

  7. [7]

    3d-future: 3d fur- niture shape with texture.IJCV, 129(12):3313–3337, 2021

    Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d fur- niture shape with texture.IJCV, 129(12):3313–3337, 2021. 3

    2021

  8. [8]

    Sparseflex: High-resolution and arbitrary-topology 3d shape modeling.arXiv preprint arXiv:2503.21732, 2025

    Xianglong He, Zi-Xin Zou, Chia-Hao Chen, Yuan-Chen Guo, Ding Liang, Chun Yuan, Wanli Ouyang, Yan-Pei Cao, and Yangguang Li. Sparseflex: High-resolution and arbitrary-topology 3d shape modeling.arXiv preprint arXiv:2503.21732, 2025. 2

    Pith/arXiv arXiv 2025

  9. [9]

    Lrm: Large reconstruction model for single image to 3d.arXiv preprint arXiv:2311.04400, 2023

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d.arXiv preprint arXiv:2311.04400, 2023. 2

    Pith/arXiv arXiv 2023

  10. [10]

    Sat2city: 3d city generation from a single satellite im- age with cascaded latent diffusion

    Tongyan Hua, Lutao Jiang, Ying-Cong Chen, and Wufan Zhao. Sat2city: 3d city generation from a single satellite im- age with cascaded latent diffusion. InICCV, pages 27978– 27988, 2025. 1, 2

    2025

  11. [11]

    City3d: Large-scale building reconstruction from airborne lidar point clouds.Remote Sensing, 14(9):2254, 2022

    Jin Huang, Jantien Stoter, Ravi Peters, and Liangliang Nan. City3d: Large-scale building reconstruction from airborne lidar point clouds.Remote Sensing, 14(9):2254, 2022. 3

    2022

  12. [12]

    Sat-skylines: 3d building generation from satellite imagery and coarse geometric pri- ors.arXiv preprint arXiv:2508.18531, 2025

    Zhangyu Jin and Andrew Feng. Sat-skylines: 3d building generation from satellite imagery and coarse geometric pri- ors.arXiv preprint arXiv:2508.18531, 2025. 4

    Pith/arXiv arXiv 2025

  13. [13]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1,

  14. [14]

    Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal naviga- tion

    Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal naviga- tion. InCVPR, pages 16384–16393, 2024. 3

    2024

  15. [15]

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dock- horn, Jack English, Zion English, Patrick Esser, et al. Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space.arXiv preprint arXiv:2506.15742,

    Pith/arXiv arXiv

  16. [16]

    Hunyuan3d 2.5: Towards high- fidelity 3d assets generation with ultimate details.arXiv preprint arXiv:2506.16504, 2025

    Zeqiang Lai, Yunfei Zhao, Haolin Liu, Zibo Zhao, Qingxi- ang Lin, Huiwen Shi, Xianghui Yang, Mingxin Yang, Shuhui Yang, Yifei Feng, et al. Hunyuan3d 2.5: Towards high- fidelity 3d assets generation with ultimate details.arXiv preprint arXiv:2506.16504, 2025. 2, 4

    Pith/arXiv arXiv 2025

  17. [17]

    Unleashing vecset diffusion model for fast shape generation.arXiv preprint arXiv:2503.16302,

    Zeqiang Lai, Yunfei Zhao, Zibo Zhao, Haolin Liu, Fuyun Wang, Huiwen Shi, Xianghui Yang, Qingxiang Lin, Jingwei Huang, Yuhong Liu, et al. Unleashing vecset diffusion model for fast shape generation.arXiv preprint arXiv:2503.16302,

    Pith/arXiv arXiv

  18. [18]

    Nuiscene: Exploring efficient generation of unbounded out- door scenes.arXiv preprint arXiv:2503.16375, 2025

    Han-Hung Lee, Qinghong Han, and Angel X Chang. Nuiscene: Exploring efficient generation of unbounded out- door scenes.arXiv preprint arXiv:2503.16375, 2025. 3

    Pith/arXiv arXiv 2025

  19. [19]

    Crafts- man3d: High-fidelity mesh generation with 3d native gen- eration and interactive geometry refiner.arXiv preprint arXiv:2405.14979, 2024

    Weiyu Li, Jiarui Liu, Hongyu Yan, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. Crafts- man3d: High-fidelity mesh generation with 3d native gen- eration and interactive geometry refiner.arXiv preprint arXiv:2405.14979, 2024. 2

    Pith/arXiv arXiv 2024

  20. [20]

    Step1x-3d: Towards high-fidelity and con- trollable generation of textured 3d assets.arXiv preprint arXiv:2505.07747, 2025

    Weiyu Li, Xuanyang Zhang, Zheng Sun, Di Qi, Hao Li, Wei Cheng, Weiwei Cai, Shihao Wu, Jiarui Liu, Zihao Wang, et al. Step1x-3d: Towards high-fidelity and con- trollable generation of textured 3d assets.arXiv preprint arXiv:2505.07747, 2025. 6

    Pith/arXiv arXiv 2025

  21. [21]

    Triposg: High-fidelity 3d shape synthesis using large-scale rectified flow models.arXiv preprint arXiv:2502.06608, 2025

    Yangguang Li, Zi-Xin Zou, Zexiang Liu, Dehu Wang, Yuan Liang, Zhipeng Yu, Xingchao Liu, Yuan-Chen Guo, Ding Liang, Wanli Ouyang, et al. Triposg: High-fidelity 3d shape synthesis using large-scale rectified flow models.arXiv preprint arXiv:2502.06608, 2025. 2

    Pith/arXiv arXiv 2025

  22. [22]

    Sat2scene: 3d urban scene generation from satellite images with diffusion

    Zuoyue Li, Zhenqiang Li, Zhaopeng Cui, Marc Pollefeys, and Martin R Oswald. Sat2scene: 3d urban scene generation from satellite images with diffusion. InCVPR, pages 7141– 7150, 2024. 1, 2

    2024

  23. [23]

    Capturing, reconstructing, and simulating: the urbanscene3d dataset

    Liqiang Lin, Yilin Liu, Yue Hu, Xingguang Yan, Ke Xie, and Hui Huang. Capturing, reconstructing, and simulating: the urbanscene3d dataset. InEuropean Conference on Computer Vision, pages 93–109. Springer, 2022. 3

    2022

  24. [24]

    One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion

    Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Ji- ayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. InCVPR, pages 10072–10083, 2024. 2 9

    2024

  25. [25]

    Zero-1-to-3: Zero-shot one image to 3d object

    Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tok- makov, Sergey Zakharov, and Carl V ondrick. Zero-1-to-3: Zero-shot one image to 3d object. InICCV, pages 9298– 9309, 2023. 1, 2

    2023

  26. [26]

    Earthcrafter: Scalable 3d earth generation via dual-sparse latent diffusion.arXiv preprint arXiv:2507.16535, 2025

    Shang Liu, Chenjie Cao, Chaohui Yu, Wen Qian, Jing Wang, and Fan Wang. Earthcrafter: Scalable 3d earth generation via dual-sparse latent diffusion.arXiv preprint arXiv:2507.16535, 2025. 1, 2

    Pith/arXiv arXiv 2025

  27. [27]

    World- craft: Photo-realistic 3d world creation and customization via llm agents.arXiv preprint arXiv:2502.15601, 2025

    Xinhang Liu, Chi-Keung Tang, and Yu-Wing Tai. World- craft: Photo-realistic 3d world creation and customization via llm agents.arXiv preprint arXiv:2502.15601, 2025. 1

    Pith/arXiv arXiv 2025

  28. [28]

    Urban architect: Steerable 3d urban scene generation with layout prior.arXiv preprint arXiv:2404.06780, 2024

    Fan Lu, Kwan-Yee Lin, Yan Xu, Hongsheng Li, Guang Chen, and Changjun Jiang. Urban architect: Steerable 3d urban scene generation with layout prior.arXiv preprint arXiv:2404.06780, 2024. 1

    Pith/arXiv arXiv 2024

  29. [29]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021. 2

    2021

  30. [30]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 5

    Pith/arXiv arXiv 2023

  31. [31]

    Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022. 1, 2

    Pith/arXiv arXiv 2022

  32. [32]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. InICML, pages 8748–8763. PmLR, 2021. 6

    2021

  33. [33]

    Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies

    Xuanchi Ren, Jiahui Huang, Xiaohui Zeng, Ken Museth, Sanja Fidler, and Francis Williams. Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. In CVPR, pages 4209–4219, 2024. 2

    2024

  34. [34]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InCVPR, pages 10684– 10695, 2022. 1

    2022

  35. [35]

    Seed3d 1.0: From images to high-fidelity simulation-ready 3d assets

    ByteDance Seed. Seed3d 1.0: From images to high-fidelity simulation-ready 3d assets. 8

  36. [36]

    Ur- banworld: An urban world model for 3d city generation

    Yu Shang, Yuming Lin, Yu Zheng, Hangyu Fan, Jingtao Ding, Jie Feng, Jiansheng Chen, Li Tian, and Yong Li. Ur- banworld: An urban world model for 3d city generation. arXiv preprint arXiv:2407.11965, 2024. 1, 2

    Pith/arXiv arXiv 2024

  37. [37]

    Dinov3.arXiv preprint arXiv:2508.10104, 2025

    Oriane Sim ´eoni, Huy V V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha ¨el Ramamonjisoa, et al. Dinov3.arXiv preprint arXiv:2508.10104, 2025. 5

    Pith/arXiv arXiv 2025

  38. [38]

    Lgm: Large multi-view gaus- sian model for high-resolution 3d content creation

    Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaus- sian model for high-resolution 3d content creation. InECCV, pages 1–18. Springer, 2024. 2

    2024

  39. [39]

    Hunyuan3d 2.1: From images to high-fidelity 3d assets with production-ready pbr material,

    Tencent Hunyuan3D Team. Hunyuan3d 2.1: From images to high-fidelity 3d assets with production-ready pbr material,

  40. [40]

    Hunyuan3d 2.0: Scaling diffu- sion models for high resolution textured 3d assets generation,

    Tencent Hunyuan3D Team. Hunyuan3d 2.0: Scaling diffu- sion models for high resolution textured 3d assets generation,

  41. [41]

    Building3d: A urban-scale dataset and benchmarks for learn- ing roof structures from point clouds

    Ruisheng Wang, Shangfeng Huang, and Hongxin Yang. Building3d: A urban-scale dataset and benchmarks for learn- ing roof structures from point clouds. InICCV, pages 20076– 20086, 2023. 3

    2023

  42. [42]

    Realcity3d: A large-scale georeferenced 3d shape dataset of real-world cities

    Congcong Wen, Wenyu Han, Lazarus Chok, Yan Liang Tan, Sheung Lung Chan, Hang Zhao, and Chen Feng. Realcity3d: A large-scale georeferenced 3d shape dataset of real-world cities. 3

  43. [43]

    Structured 3d latents for scalable and versatile 3d gen- eration

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d gen- eration. InCVPR, pages 21469–21480, 2025. 1, 2, 3, 4, 5, 6

    2025

  44. [44]

    Sana: Efficient high-resolution image syn- thesis with linear diffusion transformers.arXiv preprint arXiv:2410.10629, 2024

    Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image syn- thesis with linear diffusion transformers.arXiv preprint arXiv:2410.10629, 2024. 1

    Pith/arXiv arXiv 2024

  45. [45]

    Citydreamer: Compositional generative model of unbounded 3d cities

    Haozhe Xie, Zhaoxi Chen, Fangzhou Hong, and Ziwei Liu. Citydreamer: Compositional generative model of unbounded 3d cities. InCVPR, pages 9666–9675, 2024. 1, 8

    2024

  46. [46]

    Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models.arXiv preprint arXiv:2404.07191,

    Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models.arXiv preprint arXiv:2404.07191,

    Pith/arXiv arXiv

  47. [47]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025. 6

    Pith/arXiv arXiv 2025

  48. [48]

    Hunyuan3d 1.0: A unified framework for text-to-3d and image-to-3d generation.arXiv preprint arXiv:2411.02293, 2024

    Xianghui Yang, Huiwen Shi, Bowen Zhang, Fan Yang, Ji- acheng Wang, Hongxu Zhao, Xinhai Liu, Xinzhou Wang, Qingxiang Lin, Jiaao Yu, et al. Hunyuan3d 1.0: A unified framework for text-to-3d and image-to-3d generation.arXiv preprint arXiv:2411.02293, 2024. 2

    Pith/arXiv arXiv 2024

  49. [49]

    3dshape2vecset: A 3d shape representation for neu- ral fields and generative diffusion models.TOG, 42(4):1–16,

    Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neu- ral fields and generative diffusion models.TOG, 42(4):1–16,

  50. [50]

    Gs-lrm: Large recon- struction model for 3d gaussian splatting

    Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large recon- struction model for 3d gaussian splatting. InECCV, pages 1–19. Springer, 2024. 2

    2024

  51. [51]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836–3847, 2023. 1

    2023

  52. [52]

    Clay: A controllable large-scale generative model for creating high-quality 3d assets.TOG, 43(4):1–20, 2024

    Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets.TOG, 43(4):1–20, 2024. 2

    2024

  53. [53]

    Cityx: Controllable procedural content generation for unbounded 3d cities.arXiv preprint arXiv:2407.17572,

    Shougao Zhang, Mengqi Zhou, Yuxi Wang, Chuanchen Luo, Rongyu Wang, Yiwei Li, Zhaoxiang Zhang, and Junran Peng. Cityx: Controllable procedural content generation for unbounded 3d cities.arXiv preprint arXiv:2407.17572,

    Pith/arXiv arXiv