OmniTex-Gen: Texturing LOD-1 Building Models via Panorama Synthesis from Virtual Street-View Depth
Xuming Zhang, Han Hu*, Liupeng Su, Junfan Wang, Yuhao Ye, Qing Zhu
Faculty of Geosciences and Engineering, Southwest Jiaotong University, Chengdu 611756, China
* Corresponding author: han.hu@swjtu.edu.cn
preprint· 2026
Abstract
Satellite imagery has become a major data source for large-scale reconstruction of Level of Detail 1 (LOD-1) building models. However, satellite imagery provides limited views of façades, so these models typically use textures from generic libraries, limiting their visual realism. We present a framework that projects LOD-1 building models into virtual street-view depth maps to guide panorama synthesis, in order to generate realistic façade textures consistent with the building geometry. Specifically, a latent diffusion model uses these maps through ControlNet to constrain scene geometry. In addition, city-based textual conditioning provides shared streetscape cues, while satellite orthophoto features supply local appearance information through IP-Adapter.
For training and evaluation, we construct OmniCityTex, containing 29,881 samples from 18 cities that pair LOD-1 building models, satellite orthophotos, virtual depth maps, and cleaned street-view panoramas. Experimental evaluations reveal that our method reduces the DINO-based structural distance and FID by 6.2% and 39.5%, respectively, and increases building-class intersection over union by 27.1% relative to the strongest baseline for each metric. Our model achieves lower perceptual distances and higher semantic agreement than the baselines when evaluating from four unseen cities.
Contributions
- OmniCityTex dataset. We introduce a dataset of 29,881 samples from 18 cities across 15 countries. Each sample pairs an LOD-1 building model with a satellite orthophoto, a cleaned street-view panorama, and a corresponding virtual street-view depth map.
- Geometry- and appearance-conditioned panorama synthesis. We project LOD-1 building footprints and heights into virtual street-view depth maps to constrain building positions and shapes during panorama synthesis. City-based text prompts and satellite orthophoto features complement this geometric guidance by guiding façade appearance.
OmniCityTex
We construct OmniCityTex, a dataset for generating textures for LOD-1 building models, containing 29,881 samples from 18 cities across 15 countries. Each sample pairs an LOD-1 building model and a satellite orthophoto with a cleaned street-view panorama and a virtual street-view depth map from the same location.
- LOD-1 building model.
- We use the PlaceMaker plugin for SketchUp to construct LOD-1 building models from OpenStreetMap footprints and building heights.
- Satellite orthophoto.
- We use each panorama center to obtain the corresponding satellite orthophoto. The orthophoto covers the same area as the local OSM grid.
- Virtual street-view depth map.
- We project the OSM building footprints and heights into virtual depth maps aligned with the panoramas. Each pixel records the distance from the camera to the visible surface. We use our own web-based review tool to screen panorama-depth pairs and fine-tune their horizontal alignment.
Cleaned street-view panorama Loading…
Virtual street-view depth map Loading…Drag to align the depth with the panorama. - Cleaned street-view panorama.
- We use a large vision-language model to remove foreground objects from the panoramas, including vehicles, pedestrians, and vegetation. The editing prompt instructs the model to retain the camera viewpoint, panoramic geometry, building outlines, and visible façade appearance.
Loading…
Drag the divider to compare before and after cleaning, or use the arrow keys.
Comparison with State-of-the-Art Methods
Scene
Comparison method
Effect of city-based textual conditioning
Effect of visual conditioning
Drag the divider to compare, or use the arrow keys.
Generalization Analysis on Unseen Cities
Scene
Comparison method
BibTeX
@misc{zhang2026omnitexgen,
title = {OmniTex-Gen: Texturing LOD-1 Building Models via Panorama Synthesis from Virtual Street-View Depth},
author = {Zhang, Xuming and Hu, Han and Su, Liupeng and Wang, Junfan and Ye, Yuhao and Zhu, Qing},
note = {Preprint},
year = {2026}
}Acknowledgements
This study was supported in part by the National Natural Science Foundation of China (Project No. 42230102, U25A20772) and the Natural Science Foundation of Sichuan Province under Grant 2026NSFSCZY0054.