Many urban applications need high-resolution elevation data, but fine-resolution DSMs are expensive to acquire. This work combines two widely available inputs: coarse 5 m DSMs and high-resolution RGB imagery. A diffusion-based generator is conditioned on the RGB image and denoises the elevation map, transferring building outlines and roof structure visible only in the image into the final DSM.
The method was evaluated on cities in Central Europe and produced high-quality DSMs with improved structural detail and accurate geometry compared with conventional interpolation or filtering. To address the slow iterative sampling typical of diffusion models, the approach uses an end-to-end single-step finetuning strategy that preserves DSM quality at a fraction of the cost. Because the generator is pretrained on billions of images, it generalizes across cities and can be sampled multiple times to quantify uncertainty.