Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

EarthCrafter: Scalable 3D Earth Generation via Dual-Sparse Latent Diffusion

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read EarthCrafter: a dual-sparse latent diffusion framework trained on 50,028 annotated 600-meter aerial scenes can generate plausible 3D terrain from a semantic map, a single image, or no condition at all.

desk verdict Aerial-Earth3D is a real contribution and EarthCrafter is a serious engineering effort, but the paper's 'substantially better' claim is not supported by the numbers as presented. read the letter →

arxiv 2507.16535 v2 pith:Z5FAJTNZ submitted 2025-07-22 cs.CV cs.AI

classification cs.CVcs.AI
keywords 3DEarthgenerationlatentdiffusionflowmatchingsparsevoxelrepresentation2DGaussiansplattingaerialdatasetsemantic-guidedgeographic-scalescenes
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

EarthCrafter aims to prove that 3D generation can be scaled from single objects and small city patches to geographic extents—hundreds of meters of real terrain at a time—without collapsing into incoherent geometry or blurry texture. To make that scale tractable, the paper compresses the two expensive parts of a 3D scene separately: a structural variational autoencoder squeezes the voxel grid of geometry to 1/256 of its volume, while a texture autoencoder keeps the voxel resolution but compresses feature channels, decoding to 2D Gaussian splats. On top of these latents, two flow-matching models accept a semantic map, a posed RGB(D) image, or no condition at all, and the paper reports that this performs substantially better than prior urban generators at extremely large scales. The paper also contributes Aerial-Earth3D, a 50,028-scene, 45-million-frame dataset of curated U.S. terrain annotated with depth, normals, semantics, and camera poses, arguing that this data is what makes geographically plausible generation possible. If the system is right, unconditional and semantic-guided 3D Earth synthesis becomes a practical starting point for simulation, urban planning, and virtual environment production.

What carries the argument

The load-bearing mechanism is the dual-sparse latent decomposition: two VAEs compress geometry and appearance into separate small latents so that the diffusion models run on cheap representations instead of full Earth-scale volumes. For geometry, StructVAE downsamples via stride-2 3D sparse convolutions and upsamples through novel Pseudo-Sparse to Sparse (PSS) blocks, where a sparse pixel shuffle creates pseudo-sparse voxels, a classification head decides which are valid, and the rest are pruned; the paper reports that this classification scaffold is what makes accurate sparse upsampling possible, reaching 97.1% structural accuracy. For appearance, TexVAE deliberately avoids spatial compression—the paper's ablation shows that spatially compressing texture features degrades recovery—and instead compresses channels, consuming fine-grained low-channel image-VAE features (66 channels, about 31M storage per scene rather than 471M for large-channel alternatives) and decoding to voxelized 2D Gaussian Splats with 16 primitives per voxel. Two flow-matching generators sit on these latents: coarse-to-fine StructFM separates voxel classification from latent prediction because the paper finds that fusing them in one dense model hurts both tasks, and TexFM uses additive condition injection rather than cross-attention because cross-attention showed little correspondence between condition pixels and generated texture.

What would settle it

Regenerate a subset of the dataset's footprints with EarthCrafter and compare the outputs voxel-by-voxel against independent airborne LiDAR or high-resolution satellite stereo meshes of the same sites, using the paper's L/8 and L-level mIoU metrics. If the scores against LiDAR-derived geometry drop sharply while scores against the paper's InstantNGP meshes stay high, the learned geometry encodes reconstruction artifacts rather than real terrain; if the scores hold, the InstantNGP ground truth was not the limiting factor.

Watch

Extended reading notes

Core claim

The central claim is that large-scale 3D Earth generation can be split into two much smaller generative problems without losing geographic plausibility. EarthCrafter trains a structural VAE (StructVAE) that compresses a full voxel coordinate grid $V_c \in \mathbb{R}^{L^3 \times 3}$ into a structural latent $S_{La} \in \mathbb{R}^{(L/8)^3 \times 32}$ using Pseudo-Sparse to Sparse (PSS) blocks that upsample pseudo-sparse voxels and classify away invalid ones, and a texture VAE (TexVAE) that sends 66-channel voxelized features—hierarchical image-VAE features plus RGB and normals—into an 8-channel textural latent without spatial downsampling, decoding to 16 Gaussian primitives per voxel. Two condition-aware flow-matching models act on these latents: StructFM works coarse-to-fine, first classifying activated voxels at $L/8$ resolution and then refining coordinates while predicting structural latents, while TexFM generates textural latents on the decoded voxels with a U-Net that mixes sparse Swin attention at fine scales with full attention at coarse scales. A single forward pass at $L=256$ therefore produces a 146-square-meter scene, and a sliding-window inpainting scheme on a large semantic map extends this to 412-square-meter regions; the prior that keeps unconditioned terrain geographically plausible comes from training on the 50,028-scene dataset. Evidence consists of architectural ablations (voxel classification accuracy of 97.1%, PSNR and LPIPS texture gains, and a +3.3 mIoU gain from the coarse-to-fine separation) plus qualitative comparisons against prior urban generators, with the paper concluding that the gains are largest specifically in the extremely-large-scale regime.

Load-bearing premise

The pipeline assumes that InstantNGP reconstructions from a limited set of Google Earth aerial views are accurate enough to serve as ground-truth geometry for both training and evaluation, so any systematic reconstruction error—holes at tree corners, distortions on building tops and facades—is baked into what the models learn and how they are judged.

Editorial extensions

If this is right

  • A single forward pass yields a 146-square-meter 3D scene, and applying the same generative model in a sliding window over a large semantic map extends this to 412-square-meter regions and, by iteration, arbitrarily large areas.
  • The same trained framework covers semantic-guided urban layout generation, single-image RGBD-conditioned generation, and fully unconditional terrain synthesis, because the flow models are trained on mixed conditions including an explicit empty condition.
  • Aerial-Earth3D, with 50,028 scenes and 45 million multi-view frames together with depth, normal, semantic, and pose annotations, is positioned as the substrate that later 3D and 2D tasks—semantic segmentation, building height estimation, outline prediction, outdoor reconstruction—can build on.
  • Splitting structure from texture lets each part be compressed to its own latent scale (1/256 volume for geometry, channel-compressed full-resolution for texture), which the paper credits for making extreme-scale generation tractable where dense object-level frameworks are not.
  • The coarse-to-fine StructFM results imply that voxel classification and latent feature prediction conflict when forced into one dense model; separating them improves voxel classification by 3.3 mIoU at the coarse level and also improves fine-level voxel quality.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A practical extension the paper leaves implicit is text- or layout-conditioned planning: reversing the semantic-prior direction, so that a land-use sketch is first synthesized from a text prompt and then fed to EarthCrafter, would turn the framework into a testable urban-planning sandbox.
  • Given the paper's own limitation about facade and tree-corner distortion, incorporating street-level or oblique imagery into the reconstruction is the most direct attack on the weakest part of the ground truth; a testable prediction is that voxel IoU on building side surfaces will be the dominant bottleneck in any independent geometric evaluation.
  • All comparisons at the 412-square-meter scale are qualitative; a quantitative validation against a real aerial mosaic of the same footprint, or a perceptual study comparing generated and real renders, would be the natural next step and would determine how much of the claimed large-scale advantage is photorealism rather than geographic fidelity.
  • The dataset's semantic distribution is dominated by woodland, grassland, building, pavement, and road, so the diversity gains from the paper's height-weighted sampling are most likely concentrated in the long tail of rare classes such as water, bridges, vehicles, and helipads rather than in common ground cover.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. EarthCrafter proposes a two-stage framework for large-scale 3D Earth generation. It introduces Aerial-Earth3D, a dataset of 50,028 curated 600m x 600m scenes with 45M multi-view Google Earth frames, plus dual sparse VAEs that compress geometry voxels and 2D Gaussian splatting features into compact latents, and flow-matching models that generate structure and texture under semantic, RGBD, or empty conditioning. The paper also contributes a coarse-to-fine structural flow model, a pseudo-sparse-to-sparse upsampling block, and an infinite-scene extension via sliding-window inpainting. The central claim is that EarthCrafter performs substantially better in extremely large-scale generation than prior work.

Significance. The Aerial-Earth3D dataset, if released, is a substantial resource: it is an order of magnitude larger than existing aerial 3D scene datasets and includes multi-view images, depth, normals, semantics, and camera poses. The architectural ideas, especially the PSS block for sparse upsampling, the coarse-to-fine StructFM, and the use of low-channel Flux-VAE features for texture, are well motivated and are supported by internal ablations. However, the comparative and generalization claims are not yet supported by the evaluation: the only cross-method comparison is qualitative, and the headline quantitative numbers are reported on training data. This is a promising systems contribution whose central claim needs stronger evidence before acceptance.

major comments (4)
  1. [Qualitative Comparison / Figure 6] The abstract and introduction claim that EarthCrafter performs substantially better in extremely large-scale generation, but the only cross-method evidence is the qualitative comparison in Figure 6, which the authors explicitly characterize as non-quantitative because no research shares similar settings. This is load-bearing: a comparative superiority claim needs either quantitatively comparable baselines on a shared protocol (for example, running CityDreamer and SceneDreamer on the same semantic maps and computing FID, KVD, or geometry statistics on held-out scenes) or a revised claim that the contribution is the first scalable framework rather than substantially better. As written, the central claim is not established by the experiments.
  2. [Tables 3, 4, 16] The headline quantitative numbers are computed on training data. Table 3 reports StructVAE accuracy on ablation and train data, with the 97.1% figure coming from the global training rows; Table 4 is captioned 'Results of StructFlows on train data under image condition'; Table 16 likewise reports train-data mIoU. A validation split of 3,068 samples (2,801 for the flow models) is described in Data Preparation, but no validation numbers are reported for these models. Reporting only training-set reconstruction and generation metrics does not demonstrate generalization and can reflect memorization; the authors should report the corresponding validation metrics or explicitly justify why train-data metrics are the appropriate evidence.
  3. [Data Preparation; Infinite Scene Generation] The evaluation protocol is not held out at the scene level. The train/validation split in Data Preparation is performed by height-stratified sampling from the same 450k voxel-feature pool, which is generated by sliding-crop from the 50k scenes; thus crops from the same 600m scene can appear in both training and validation. Moreover, the note in Infinite Scene Generation states that the large semantic map used for the flagship Figure 7 is taken from a validation patch whose source scene mesh has overlap with training patches. This is train/evaluation leakage and undermines the claim that Figure 7 demonstrates scalable generation of novel large scenes. Please re-run the infinite-scene demonstration and any quantitative extension metrics on a geographically disjoint held-out region and report the overlap statistics.
  4. [Achievement of Annotated Mesh; Limitations] The ground-truth geometry for both training and evaluation is the InstantNGP reconstruction from a limited set of Google Earth aerial views. The Limitations section admits noticeable holes and distortions at the corners of trees and the upper portions of tall buildings, and that side surfaces such as facades are lower fidelity. Because all reported structure and texture metrics compare against this voxelized proxy, systematic reconstruction errors could be learned by the models and scored as geographic plausibility even when generated scenes deviate from real terrain. This does not invalidate the system, but it is a correctness risk for absolute accuracy claims such as the 97.1% StructVAE accuracy. A concrete mitigation would be to validate the proxy on a small set of scenes with higher-quality reconstructions (more views or LiDAR-derived mesh) and report how the metrics change, or at least discuss the direction of the bias.
minor comments (6)
  1. [StructVAE ablations] The section header contains a typo: 'StruceV AE' should be 'StructVAE'.
  2. [Table 12] The baseline row of the TexFM ablation table lacks explicit markers for the C-Block, Layers, Channels, and Cross columns, making it difficult to identify which configuration is being compared; please align the table with clear symbols for all configurations.
  3. [Data Preparation] The text states that each voxel represents an area of 0.56m^3, calculated as 200/360; 200/360 is a side length in meters, so the units should be corrected to a voxel side length or volume.
  4. [Figure 6] The caption refers only to CityDreamer, but the surrounding text discusses SceneDreamer as well; please ensure the figure labels and caption match the methods discussed.
  5. [Limitations] The first sentence of the Limitations section is a run-on ('While EarthCrafter exhibits enhanced and diverse capabilities... However it does have...'); please rephrase.
  6. [References] The references for Google Earth Studio, GoogleMap, and PyMeshFix are listed without full citation details; please complete these entries.

Circularity Check

0 steps flagged · score 0.0 of 10

No definitional circularity: the claimed generation pipeline is self-contained; the principal weaknesses are evaluation leakage and train-set reporting, which are evidence issues, not circular reasoning.

full rationale

The paper's central derivation — StructFM samples a structural latent from noise, StructVAE decodes it to voxel coordinates, TexFM samples a textural latent conditioned on those coordinates, and TexVAE decodes it to 2DGS — is a standard conditional flow-matching generation chain. Every reported quantity (StructVAE accuracy, mIoU, PSNR, FID, LPIPS) is computed against an external voxelized proxy (InstantNGP-reconstructed meshes) rather than being defined in terms of the model's own outputs; none of the loss equations (Ltexvae, Lstructvae, or the flow-matching objective) reduces a predicted quantity to its own input. The self-citation for score aggregation (VCD-Texture, Liu et al. 2024b) is a non-load-bearing design choice with an independent external publication. The manuscript itself flags the two genuine validity threats: Table 3, Table 4, and Table 16 are explicitly reported on train data, and the Infinite Scene Generation note admits that the large semantic map is taken from a validation patch's source scene mesh that overlaps with training patches. These are evaluation-leakage and evidence-strength problems that undermine the comparative 'substantially better' claim, but they are not circularity: the model's outputs are not equivalent by construction to its training inputs or to any cited self-result. No circular step is exhibited, so the circularity score is 0.

Assumptions & free parameters 10 free parameters · 6 assumptions · 0 invented entities

The system's output depends on a chain of assumptions: Google Earth imagery and Studio rendering are usable at scale, InstantNGP meshes are accurate ground truth, AIE-SEG semantics are correct, and FLUX-VAE features are adequate texture priors. The free parameters are largely training and inference hyperparameters, many of which are reported, but several (alpha, sigma_s) are not. No new physical entities are postulated, and the paper introduces no external falsifiable predictions separate from the dataset and trained models.

free parameters (10)
  • TexVAE loss weights = lambda_l1=20, lambda_ssim=2, lambda_vgg=1.4, lambda_alex=0.6
    Hand-tuned weights in the combined L1/SSIM/LPIPS loss; directly determine texture reconstruction quality.
  • StructVAE classification loss weights = lambda_0=lambda_1=lambda_2=5
    Cross-entropy weights for the three PSS upsampling levels; chosen by hand and applied uniformly.
  • Latent voxel activation thresholds = SLa > 0.3 and >50% of channels above threshold
    Post-training conversion rule for deciding which predicted structural latents become valid voxels; central to the coarse-to-fine geometry output and not derived from data.
  • Flow matching scheduler shift = 3.0
    Noise schedule shift for DiscreteScheduler in both FM models; chosen for stability, no sensitivity analysis shown.
  • CFG strength and sampling steps = 3.0 and 25 steps
    Inference hyperparameters for all generated results; quality and diversity depend on them.
  • Score aggregation exponents = tau_s=3, tau_d=3, Z_far=2.0
    Manually set exponents and max depth in the multi-view feature fusion formula; shape all mesh vertex features.
  • Unbalanced sampling exponent alpha = not reported
    Equation (2) defines z_i=(min(Z_i,200)/10)^alpha to counter class imbalance, but alpha is never specified in the paper.
  • Normal drop threshold sigma_s = not reported
    Threshold for voxel dropping based on normal cosine similarity in the StructFM augmentation; the paper does not give its value.
  • Voxel side length = 0.56 m (200 m / 360 voxels)
    Physical size of each training voxel; fixes the scale of all generated scenes and the 146m/412m generation figures.
  • Dummy depth for semantic/empty conditions = z=128
    Semantic and empty conditions are projected into 3D using a fixed dummy depth of 128, which imposes assumed scene height.
assumptions (6)
  • domain assumption Google Earth imagery and Google Earth Studio rendering are available and permissible for dataset construction and release
    The entire Aerial-Earth3D dataset is captured from Google Earth; redistribution and derived data licensing is assumed but not discussed.
  • domain assumption InstantNGP reconstructions from about 72 aerial views per scene produce sufficiently accurate 3D meshes to serve as ground truth geometry
    Meshes are reconstructed with InstantNGP and then voxelized as structure ground truth; the Limitations section admits holes and distortions in trees and building facades.
  • domain assumption AIE-SEG semantic predictions from aerial views are reliable enough to serve as semantic ground truth
    Semantic maps are generated by AIE-SEG (trained on proprietary aerial data), not manually verified; the 25-class distribution is used to condition generation.
  • domain assumption FLUX-VAE low-channel features preserve texture information sufficient to train a 2DGS texture decoder
    The paper abandons DINOv2 1024-d features in favor of FLUX 16-d hierarchical features; this choice underlies the TexVAE design and is validated only by internal ablation.
  • standard math Flow matching objective from Trellis transfers to the dual-sparse setting
    TexFM and StructFM are trained with the same conditional flow matching objective as Trellis; correctness of that objective is assumed.
  • domain assumption Sparse voxel geometry with 0.56m voxel resolution captures the structural detail needed for geographic-scale generation
    Voxelization discards sub-voxel details; the claimed fidelity depends on this resolution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EarthCrafter: Scalable 3D Earth Generation via Dual-Sparse Latent Diffusion." pith.science (2026). https://pith.science/paper/Z5FAJTNZ

@misc{pith2026250716535,
  author       = {Pith},
  title        = {Pith review of: EarthCrafter: Scalable 3D Earth Generation via Dual-Sparse Latent Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z5FAJTNZ}},
  note         = {Machine review of arXiv:2507.16535}
}
read the original abstract

Despite the remarkable developments achieved by recent 3D generation works, scaling these methods to geographic extents, such as modeling thousands of square kilometers of Earth's surface, remains an open challenge. We address this through a dual innovation in data infrastructure and model architecture. First, we introduce Aerial-Earth3D, the largest 3D aerial dataset to date, consisting of 50k curated scenes (each measuring 600m x 600m) captured across the U.S. mainland, comprising 45M multi-view Google Earth frames. Each scene provides pose-annotated multi-view images, depth maps, normals, semantic segmentation, and camera poses, with explicit quality control to ensure terrain diversity. Building on this foundation, we propose EarthCrafter, a tailored framework for large-scale 3D Earth generation via sparse-decoupled latent diffusion. Our architecture separates structural and textural generation: 1) Dual sparse 3D-VAEs compress high-resolution geometric voxels and textural 2D Gaussian Splats (2DGS) into compact latent spaces, largely alleviating the costly computation suffering from vast geographic scales while preserving critical information. 2) We propose condition-aware flow matching models trained on mixed inputs (semantics, images, or neither) to flexibly model latent geometry and texture features independently. Extensive experiments demonstrate that EarthCrafter performs substantially better in extremely large-scale generation. The framework further supports versatile applications, from semantic-guided urban layout generation to unconditional terrain synthesis, while maintaining geographic plausibility through our rich data priors from Aerial-Earth3D. Our project page is available at https://whiteinblue.github.io/earthcrafter/

Figures

Figures reproduced from arXiv: 2507.16535 by the authors.

Figure 1
Figure 1. EarthCrafter enjoys impressive generations conditioned on various guidance, including (a) 1-view aerial semantic and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall data pipeline of Aerial-Earth3D. In￾stantNGP is utilized to achieve source meshes, which are refined with heuristic strategies. Multi-view Flux-VAE fea￾tures and semantic maps are aggregated on meshes. Then, these featured meshes are voxelized as inputs to TexVAE. were sampled from the “Things to do” recommendations in Google Earth (GoogleMap), which yielded approximately 150,745 sites of interest across… view at source ↗
Figure 3
Figure 3. Overview of EarthCrafter. EarthCrafter separately models texture and structure in the latent space compressed by TexVAE and StructVAE as illustrated in (a) and (b), respectively. EarthCrafter also contains textural and structural flow-matching models, i.e., TexFM and StructFM, to model related latent presentations. We show the overall pipeline of EarthCrafter in (c), while dashed boxes denote optional conditions [P… view at source ↗
Figures from the paper (22 more)
Figure 4
Figure 4. Figure 4: StructVAE. (a) Overview of encoder-decoder based StructVAE. (b) Pseudo-Sparse to Sparse (PSS) block is used to upsample voxels and then classify them from pseudo￾sparse voxels into sparse outcomes as in (c). SLa ∈ R ( L 8 ) 3×cs , achieving a reduction to 1/256 of the …
Figure 5
Figure 5. Figure 5: Overview of the coarse-to-fine StructFM. (a) Con￾dition branch of StructFM, which receives optional inputs: image, semantic, or empty conditions. (b) The coarse stage is devoted to classifying activated voxels. (c) The fine-grained stage focuses on refining voxel coord…
Figure 6
Figure 6. Figure 6: Qualitative Comparison. CityDreamer* means that showing results with similar camera distances compared to EarthCrafter. 200 m × 200 m. These training meshes are then voxelized with a voxel count of L = 360 to generate voxel features Vfeat, where each voxel represents a…
Figure 7
Figure 7. Figure 7: Infinite scene (412m2 ) generation under large semantic condition map. Method Data PixShuffle C-BLock FullAttn Time Acc↑ Xcube* ablation 29.1 94.3 ablation 14.8 79.2 ablation ✓ 15.7 94.3 StructVAE ablation ✓ ✓ 16.8 95.3 ablation ✓ ✓ ✓ 19.5 94.9 global ✓ ✓ 79.3 96.6 glo…
Figure 8
Figure 8. Figure 8: Diverse scene generation under semantic condition [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Diverse texture generation under semantic condition [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Diverse texture generation without condition [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: Infinite scene (412m2 ) generation under semantic condition [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: Scene generation under RGBD image condition [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: Scene generation under semantic condition [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]
Figure 14
Figure 14. Figure 14: Scene generation without condition [PITH_FULL_IMAGE:figures/full_fig_p013_14.png]
Figure 15
Figure 15. Figure 15: StructFM architectures. (a) The condition branch [PITH_FULL_IMAGE:figures/full_fig_p014_15.png]
Figure 16
Figure 16. Figure 16: TexFM architectures. (a) The condition branch of [PITH_FULL_IMAGE:figures/full_fig_p015_16.png]
Figure 17
Figure 17. Figure 17: Results of texture reconstruction changing caused by mixed LPIPS-VGG and LPIPS-ALex loss. VGG means using [PITH_FULL_IMAGE:figures/full_fig_p016_17.png]
Figure 18
Figure 18. Figure 18: The illustration of data collection of Google Earth. [PITH_FULL_IMAGE:figures/full_fig_p016_18.png]
Figure 19
Figure 19. Figure 19: Voxel data visualizations of Aerial-Earth3D. [PITH_FULL_IMAGE:figures/full_fig_p019_19.png]
Figure 20
Figure 20. Figure 20: Segmentation comparison between Florance2 and [PITH_FULL_IMAGE:figures/full_fig_p020_20.png]
Figure 21
Figure 21. Figure 21: Pie chart of the class distribution of Aerial [PITH_FULL_IMAGE:figures/full_fig_p020_21.png]
Figure 22
Figure 22. Figure 22: Results of appearance changing caused by jagged perturbation. (b) means conditional voxels from the conditional [PITH_FULL_IMAGE:figures/full_fig_p021_22.png]
Figure 23
Figure 23. Figure 23: Visualization comparison with and without rough [PITH_FULL_IMAGE:figures/full_fig_p021_23.png]
Figure 24
Figure 24. Figure 24: Results of geometry changing caused by normal drop augmentation. (b) means conditional voxels from the conditional [PITH_FULL_IMAGE:figures/full_fig_p022_24.png]
Figure 25
Figure 25. Figure 25: Visualization of condition voxels with voxel nor [PITH_FULL_IMAGE:figures/full_fig_p022_25.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

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

    cs.CV 2025-11 conditional novelty 6.0 of 10

    A satellite-to-3D-city pipeline that generates building entities with OSM geometry priors and MLLM/T2I appearance guidance reports strong gains over existing city-generation baselines.

Reference graph

Works this paper leans on

77 extracted references · 46 canonical work pages · cited by 1 Pith paper

  1. [1]

    Cao, C.; Yu, C.; Liu, S.; Wang, F.; Xue, X.; and Fu, Y. 2024. MVGenMaster: Scaling Multi-View Generation from Any Image via 3D Priors Enhanced Diffusion Model. arXiv preprint arXiv:2411.16157

  2. [2]

    R.; Nagano, K.; Chan, M

    Chan, E. R.; Nagano, K.; Chan, M. A.; Bergman, A. W.; Park, J. J.; Levy, A.; Aittala, M.; De Mello, S.; Karras, T.; and Wetzstein, G. 2023. Generative novel view synthesis with 3d-aware diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 4217--4229

  3. [3]

    Chen, M.; Hu, Q.; Yu, Z.; Thomas, H.; Feng, A.; Hou, Y.; McCullough, K.; Ren, F.; and Soibelman, L. 2022. Stpls3d: A large-scale synthetic and real aerial photogrammetry 3d point cloud dataset. arXiv preprint arXiv:2203.09065

  4. [4]

    Chen, Y.; Zheng, C.; Xu, H.; Zhuang, B.; Vedaldi, A.; Cham, T.-J.; and Cai, J. 2024. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. arXiv preprint arXiv:2411.04924

  5. [5]

    Contributors, S. 2022. Spconv: Spatially Sparse Convolution Library. https://github.com/traveller59/spconv

  6. [6]

    Deng, J.; Chai, W.; Huang, J.; Zhao, Z.; Huang, Q.; Gao, M.; Guo, J.; Hao, S.; Hu, W.; Hwang, J.-N.; et al. 2024. Citycraft: A real crafter for 3d city generation. arXiv preprint arXiv:2406.04983

  7. [7]

    Earth-centered, E.-f. 2021. Earth-centered, Earth-fixed

  8. [8]

    East-North-Up. 2012. East-North-Up

Show all 77 references
  1. [9]

    Engstler, P.; Shtedritski, A.; Laina, I.; Rupprecht, C.; and Vedaldi, A. 2025. SynCity: Training-Free Generation of 3D Worlds. arXiv preprint arXiv:2503.16420

  2. [10]

    Esser, P.; Kulal, S.; Blattmann, A.; Entezari, R.; M \"u ller, J.; Saini, H.; Levi, Y.; Lorenz, D.; Sauer, A.; Boesel, F.; et al. 2024. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning

  3. [11]

    Fridman, R.; Abecasis, A.; Kasten, Y.; and Dekel, T. 2023. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36

  4. [12]

    P.; Barron, J

    Gao*, R.; Holynski*, A.; Henzler, P.; Brussee, A.; Martin-Brualla, R.; Srinivasan, P. P.; Barron, J. T.; and Poole*, B. 2024. CAT3D: Create Anything in 3D with Multi-View Diffusion Models. Advances in Neural Information Processing Systems

  5. [13]

    GlobalMLBF. 2022. GlobalMLBF

  6. [14]

    ???? Accessed: 2024-08-12

    GoogleEarth . ???? Accessed: 2024-08-12

  7. [15]

    ???? Accessed: 2024-08-12

    GoogleMap . ???? Accessed: 2024-08-12

  8. [16]

    o llein, L.; Bo z i c , A.; M \

    H \"o llein, L.; Bo z i c , A.; M \"u ller, N.; Novotny, D.; Tseng, H.-Y.; Richardt, C.; Zollh \"o fer, M.; and Nie ner, M. 2024. Viewdiff: 3d-consistent image generation with text-to-image models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogn...

  9. [17]

    H \"o llein, L.; Cao, A.; Owens, A.; Johnson, J.; and Nie ner, M. 2023. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 7909--7920

  10. [18]

    Hong, Y.; Zhang, K.; Gu, J.; Bi, S.; Zhou, Y.; Liu, D.; Liu, F.; Sunkavalli, K.; Bui, T.; and Tan, H. 2024. Lrm: Large reconstruction model for single image to 3d. In International Conference on Learning Representations

  11. [19]

    Hu, Q.; Yang, B.; Khalid, S.; Xiao, W.; Trigoni, N.; and Markham, A. 2021. Towards semantic segmentation of urban-scale 3D point clouds: A dataset, benchmarks and challenges. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 4977--4987

  12. [20]

    Huang, B.; Yu, Z.; Chen, A.; Geiger, A.; and Gao, S. 2024 a . 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 conference papers, 1--11

  13. [21]

    Huang, B.; Yu, Z.; Chen, A.; Geiger, A.; and Gao, S. 2024 b . 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery

  14. [22]

    Kerbl, B.; Kopanas, G.; Leimk \"u hler, T.; and Drettakis, G. 2023. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Trans. Graph., 42(4): 139--1

  15. [23]

    S.; Jeong, J.; Sohn, K.; and Shin, J

    Kim, S.; Lee, K.; Choi, J. S.; Jeong, J.; Sohn, K.; and Shin, J. 2023. Collaborative score distillation for consistent visual editing. Advances in Neural Information Processing Systems, 36: 73232--73257

  16. [24]

    Labs, B. F. 2024. FLUX. https://github.com/black-forest-labs/flux

  17. [25]

    Li, X.; Lai, Z.; Xu, L.; Qu, Y.; Cao, L.; Zhang, S.; Dai, B.; and Ji, R. 2024. Director3d: Real-world camera trajectory and 3d scene generation from text. Advances in Neural Information Processing Systems, 37: 75125--75151

  18. [26]

    Li, Y.; Jiang, L.; Xu, L.; Xiangli, Y.; Wang, Z.; Lin, D.; and Dai, B. 2023. Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 3205--3215

  19. [27]

    Liang, Y.; Yang, X.; Lin, J.; Li, H.; Xu, X.; and Chen, Y. 2024. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 6517--6526

  20. [28]

    Lin, L.; Liu, Y.; Hu, Y.; Yan, X.; Xie, K.; and Huang, H. 2022. Capturing, Reconstructing, and Simulating: the UrbanScene3D Dataset. In ECCV, 93--109

  21. [29]

    Liu, M.; Shi, R.; Chen, L.; Zhang, Z.; Xu, C.; Wei, X.; Chen, H.; Zeng, C.; Gu, J.; and Su, H. 2024 a . One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. In Proceedings of the IEEE/CVF conference on computer vision and patter...

  22. [30]

    Liu, M.; Xu, C.; Jin, H.; Chen, L.; Varma T, M.; Xu, Z.; and Su, H. 2023 a . One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36: 22226--22246

  23. [31]

    Liu, R.; Wu, R.; Van Hoorick, B.; Tokmakov, P.; Zakharov, S.; and Vondrick, C. 2023 b . Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, 9298--9309

  24. [32]

    Liu, S.; Yu, C.; Cao, C.; Qian, W.; and Wang, F. 2024 b . VCD-Texture: Variance Alignment based 3D-2D Co-Denoising for Text-Guided Texturing. In European Conference on Computer Vision, 373--389. Springer

  25. [33]

    Liu, Z.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; and Guo, B. 2021. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, 10012--10022

  26. [34]

    P.; Tancik, M.; Barron, J

    Mildenhall, B.; Srinivasan, P. P.; Tancik, M.; Barron, J. T.; Ramamoorthi, R.; and Ng, R. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1): 99--106

  27. [35]

    Models, D. E. 2001. Digital Elevation Models

  28. [36]

    M\"uller, T.; Evans, A.; Schied, C.; and Keller, A. 2022. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Trans. Graph., 41(4): 102:1--102:15

  29. [37]

    R.; Novotny, D.; Gordon, T.; Lo, W.-Y.; Johnson, J.; and Gkioxari, G

    Nikhila Ravi, J. R.; Novotny, D.; Gordon, T.; Lo, W.-Y.; Johnson, J.; and Gkioxari, G. 2020. Accelerating 3D Deep Learning with PyTorch3D. arXiv:2007.08501

  30. [38]

    OpenStreetMap. 2004. OpenStreetMap

  31. [39]

    Oquab, M.; Darcet, T.; Moutakanni, T.; Vo, H.; Szafraniec, M.; Khalidov, V.; Fernandez, P.; Haziza, D.; Massa, F.; El-Nouby, A.; et al. 2023. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193

  32. [40]

    T.; and Mildenhall, B

    Poole, B.; Jain, A.; Barron, J. T.; and Mildenhall, B. 2023. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations

  33. [41]

    ???? Accessed: 2024-08-12

    PyMeshFix . ???? Accessed: 2024-08-12

  34. [42]

    P.; and Koltun, V

    Qian-Yi Zhou, J. P.; and Koltun, V. 2018. Open3D : A Modern Library for 3D Data Processing. arXiv:1801.09847

  35. [43]

    Ren, X.; Huang, J.; Zeng, X.; Museth, K.; Fidler, S.; and Williams, F. 2024 a . Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 4209--4219

  36. [44]

    Ren, X.; Lu, Y.; Liang, H.; Wu, Z.; Ling, H.; Chen, M.; Fidler, S.; Williams, F.; and Huang, J. 2024 b . Scube: Instant large-scale scene reconstruction using voxsplats. Advances in Neural Information Processing Systems

  37. [45]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10684--10695

  38. [46]

    H.; Karras, T.; Seol, Y.; Lehtinen, J.; and Aila, T

    Samuli Laine, J. H.; Karras, T.; Seol, Y.; Lehtinen, J.; and Aila, T. 2020. Modular Primitives for High-Performance Differentiable Rendering. ACM Transactions on Graphics, 39(6)

  39. [47]

    R.; Lagun, D.; Fei-Fei, L.; Sun, D.; et al

    Sargent, K.; Li, Z.; Shah, T.; Herrmann, C.; Yu, H.-X.; Zhang, Y.; Chan, E. R.; Lagun, D.; Fei-Fei, L.; Sun, D.; et al. 2024. Zeronvs: Zero-shot 360-degree view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,...

  40. [48]

    Shang, Y.; Lin, Y.; Zheng, Y.; Fan, H.; Ding, J.; Feng, J.; Chen, J.; Tian, L.; and Li, Y. 2024. UrbanWorld: An Urban World Model for 3D City Generation. arXiv preprint arXiv:2407.11965

  41. [49]

    Shi, R.; Chen, H.; Zhang, Z.; Liu, M.; Xu, C.; Wei, X.; Chen, L.; Zeng, C.; and Su, H. 2023 a . Zero123++: a Single Image to Consistent Multi-view Diffusion Base Model. arXiv:2310.15110

  42. [50]

    Shi, Y.; Wang, P.; Ye, J.; Long, M.; Li, K.; and Yang, X. 2023 b . Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512

  43. [51]

    Shriram, J.; Trevithick, A.; Liu, L.; and Ramamoorthi, R. 2024. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199

  44. [52]

    Tang, J.; Chen, Z.; Chen, X.; Wang, T.; Zeng, G.; and Liu, Z. 2024. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, 1--18. Springer

  45. [53]

    Wang, P.; and Shi, Y. 2023. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201

  46. [54]

    Wang, Z.; Lu, C.; Wang, Y.; Bao, F.; Li, C.; Su, H.; and Zhu, J. 2023. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36

  47. [55]

    Wang Ruisheng, S. Y. H. 2023. Building3D: A urban-scale dataset and benchmarks for learning roof structures from point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 20076--20086

  48. [56]

    Williams, F.; Huang, J.; Swartz, J.; Klar, G.; Thakkar, V.; Cong, M.; Ren, X.; Li, R.; Fuji-Tsang, C.; Fidler, S.; Sifakis, E.; and Museth, K. 2024 a . fVDB: A Deep-Learning Framework for Sparse, Large Scale, and High Performance Spatial Intelligence. ACM Transactions on Graph...

  49. [57]

    Williams, F.; Huang, J.; Swartz, J.; Klar, G.; Thakkar, V.; Cong, M.; Ren, X.; Li, R.; Fuji-Tsang, C.; Fidler, S.; et al. 2024 b . fvdb: A deep-learning framework for sparse, large scale, and high performance spatial intelligence. ACM Transactions on Graphics (TOG), 43(4): 1--15

  50. [58]

    P.; Verbin, D.; Barron, J

    Wu, R.; Mildenhall, B.; Henzler, P.; Park, K.; Gao, R.; Watson, D.; Srinivasan, P. P.; Verbin, D.; Barron, J. T.; Poole, B.; et al. 2024. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition,...

  51. [59]

    Xiang, J.; Lv, Z.; Xu, S.; Deng, Y.; Wang, R.; Zhang, B.; Chen, D.; Tong, X.; and Yang, J. 2025. Structured 3D Latents for Scalable and Versatile 3D Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

  52. [60]

    Xiao, B.; Wu, H.; Xu, W.; Dai, X.; Hu, H.; Lu, Y.; Zeng, M.; Liu, C.; and Yuan, L. 2023. Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks. arXiv:2311.06242

  53. [61]

    Xie, H.; Chen, Z.; Hong, F.; and Liu, Z. 2024. Citydreamer: Compositional generative model of unbounded 3d cities. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9666--9675

  54. [62]

    Xie, H.; Chen, Z.; Hong, F.; and Liu, Z. 2025 a . CityDreamer4D : Compositional Generative Model of Unbounded 4 D Cities. arXiv 2501.08983

  55. [63]

    Xie, H.; Chen, Z.; Hong, F.; and Liu, Z. 2025 b . GaussianCity: Generative Gaussian splatting for unbounded 3D city generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

  56. [64]

    Xu, H.; Man, Y.; Yang, M.; Wu, J.; Zhang, Q.; and Wang, J. 2023 a . Analytical insight of earth: a cloud-platform of intelligent computing for geospatial big data. arXiv preprint arXiv:2312.16385

  57. [65]

    Xu, Y.; Tan, H.; Luan, F.; Bi, S.; Wang, P.; Li, J.; Shi, Z.; Sunkavalli, K.; Wetzstein, G.; Xu, Z.; et al. 2023 b . Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217

  58. [66]

    Yang, Y.; Shao, J.; Li, X.; Shen, Y.; Geiger, A.; and Liao, Y. 2024. Prometheus: 3D-Aware Latent Diffusion Models for Feed-Forward Text-to-3D Scene Generation. arXiv preprint arXiv:2412.21117

  59. [67]

    Ye, V.; Li, R.; Kerr, J.; Turkulainen, M.; Yi, B.; Pan, Z.; Seiskari, O.; Ye, J.; Hu, J.; Tancik, M.; and Kanazawa, A. 2025. gsplat: An open-source library for Gaussian splatting. Journal of Machine Learning Research, 26(34): 1--17

  60. [68]

    T.; and Wu, J

    Yu, H.-X.; Duan, H.; Herrmann, C.; Freeman, W. T.; and Wu, J. 2024 a . WonderWorld: Interactive 3D Scene Generation from a Single Image. arXiv preprint arXiv:2406.09394

  61. [69]

    T.; Cole, F.; Sun, D.; Snavely, N.; Wu, J.; et al

    Yu, H.-X.; Duan, H.; Hur, J.; Sargent, K.; Rubinstein, M.; Freeman, W. T.; Cole, F.; Sun, D.; Snavely, N.; Wu, J.; et al. 2024 b . Wonderjourney: Going from anywhere to everywhere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 6658--6667

  62. [70]

    A.; Shechtman, E.; and Wang, O

    Zhang, R.; Isola, P.; Efros, A. A.; Shechtman, E.; and Wang, O. 2018. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, 586--595

  63. [71]

    Zhang, S.; Wang, J.; Xu, Y.; Xue, N.; Rupprecht, C.; Zhou, X.; Shen, Y.; and Wetzstein, G. 2025. Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views. arXiv preprint arXiv:2502.12138

  64. [72]

    Zhao, Z.; Lai, Z.; Lin, Q.; Zhao, Y.; Liu, H.; Yang, S.; Feng, Y.; Yang, M.; Zhang, S.; Yang, X.; et al. 2025. Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation. arXiv preprint arXiv:2501.12202

  65. [73]

    Zhou, S.; Fan, Z.; Xu, D.; Chang, H.; Chari, P.; Bharadwaj, T.; You, S.; Wang, Z.; and Kadambi, A. 2024. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, 324--342. Springer

  66. [74]

    Zhu, J.; and Zhuang, P. 2023. HiFA: High-fidelity Text-to-3D Generation with Advanced Diffusion Guidance. arXiv:2305.18766

  67. [75]

    Zou, Z.-X.; Yu, Z.; Guo, Y.-C.; Li, Y.; Liang, D.; Cao, Y.-P.; and Zhang, S.-H. 2024. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition...

  68. [76]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  69. [77]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.