Conditional Latent Diffusion Models for Digital Surface Model Generation
Although open digital surface model (DSM) and digital terrain model (DTM) datasets are widely available, they are typically acquired independently, using different sensors, acquisition dates, and pro- cessing pipelines. These datasets often differ in geometry, resolution, and acquisition time, which can affect applications that rely on paired terrain and surface information. Generating DSMs directly from DTMs is a useful alternative because it creates aligned and temporally consistent data without a new surface survey. We study DTM-to-DSM generation with conditional latent diffusion. The task is to generate realistic urban surfaces from a bare-earth raster and a building-footprint mask. We compare a GAN baseline with different diffusion adaptation strategies. On LiDAR data over three Swiss cities, the diffusion models produce more realistic surfaces than the GAN baseline under distributional image metrics. ControlNet-LoRA gives the best practical balance in our analysis, with clearer qualitative controllability and stronger downstream terrain-recovery results than full denoiser fine-tuning.
| Author(s): | Debwashis Borman, Daniel Panangian, Pedro Achanccaray Diaz, Martin Eisemann, Ksenia Bittner |
|---|---|
| Published: | September 2026 |
| Type: | Article in conference proceedings |
| Book: | ECCV 2026 Workshop TwinWolrd |
| Presented at: | ECCV 2026 Workshop TwinWorld 2026 |
@inproceedings{borman2026conditional,
title = {Conditional Latent Diffusion Models for Digital Surface Model Generation},
author = {Borman, Debwashis and Panangian, Daniel and Diaz, Pedro Achanccaray and Eisemann, Martin and Bittner, Ksenia},
booktitle = {{ECCV} 2026 Workshop TwinWolrd},
month = {Sep},
year = {2026}
}
Authors
-
Debwashis Borman
Student -
Daniel Panangian
External -
Pedro Achanccaray Diaz
External -
Martin Eisemann
Director -
Ksenia Bittner
External