PaGeR: Panorama Geometry Estimation

High-Resolution Depth and Normals using Single-Step Diffusion Models

Vukasin Bozic Vukasin Bozic¹ Isidora Slavkovic Isidora Slavkovic² Dominik Narnhofer Dominik Narnhofer¹ Nando Metzger Nando Metzger¹ Denis Rozumny Denis Rozumny¹ Konrad Schindler Konrad Schindler¹ Nikolai Kalischek Nikolai Kalischek²
school ¹ETH Zurich
business ²Google
description arXiv 🤗 HF Demo code Code dataset Dataset

Abstract

Monocular geometry estimation has greatly advanced in recent years and has matured to the point where off-the-shelf, foundational models are readily available. Several such models are based on denoising diffusion, which has shown a remarkable ability to transfer learned priors from color image generation to (image-conditional) geometry reconstruction, by fine-tuning on large synthetic depth datasets. We extend diffusion-based monodepth to the more challenging task of estimating full panoramic 3D geometry.

We leverage recent advances in panorama generation and diffusion fine-tuning and introduce PaGeR (Panoramic Geometry Reconstruction): a one-step diffusion model trained directly in pixel space, which enables high-resolution prediction of panoramic depth and surface normals, with strong generalization to new, unseen environments. As part of our effort, we introduce a synthetic high-resolution dataset of indoor and outdoor scenes with associated metric depth and surface normals. Extensive experiments show that our model produces coherent, metrically accurate, sharp depth and normal maps, and outperforms prior approaches not only in domain but also in few-shot and zero-shot scenarios.

Method

Method Figure

Cubemap Projection

We project the equirectangular panorama into a cubemap representation to avoid polar distortion and leverage standard convolutional architectures.

Pixel-Space Diffusion

Unlike latent models, PaGeR operates in pixel space, preserving the high-frequency details essential for accurate normal estimation and depth discontinuities.

One-Step Inference

Through distillation, we achieve high-quality geometry estimation in a single inference step, making the method computationally efficient for high-resolution outputs.

Interactive Results

360 Depth Comparison

Interactive

Scenes

Competitor

compare Normal Comparison

Interactive

Scenes

Competitor Result
Ours Result
PaGeR (Ours) Competitor
code

view_in_ar 3D Point Cloud

Interactive

Citation

@article{bozic2026pager,
  title={PaGeR: Panorama Geometry Estimation using Single-Step Diffusion Models},
  author={Bozic, Vukasin and Slavkovic, Isidora and Narnhofer, Dominik and Metzger, Nando and Rozumny, Denis and Schindler, Konrad and Kalischek, Nikolai},
  journal={arXiv preprint},
  year={2025}
}