GeoNeXt
ECCV 2026

Video Generative Models
as Geometry Learner

Haosen Yang1 Jifei Song1 Zhensong Zhang2 Xiatian Zhu1* Jiankang Deng3*

1University of Surrey   ·   2Independent Researcher   ·   3Imperial College London

GitHub Models
Building geometry…
Building geometry…
↻ Drag to rotate · Scroll to zoom

Meshes are reconstructed directly from predicted depth and surface normals, then aligned to the MoGe-2 coordinate space for comparison.

TL;DR

  • GeoNeXt repurposes an off-the-shelf video generative model as a geometry learner for monocular depth and surface-normal estimation.
  • It reformulates geometry estimation as next-frame prediction, leveraging the video model’s temporal priors to enable cross-frame interaction between appearance and geometry.
  • By co-generating the image, depth, and normals along a shared denoising trajectory, GeoNeXt achieves strong zero-shot generalization using only 59K training samples.

How it works

Geometry as next frames

GeoNeXt fine-tuning framework with frozen VAE encoding, joint image-depth-normal noising, GeoNeXt U-Net, and reconstruction objectives
Figure 2

Overview of the GeoNeXt fine-tuning protocol. (a) We start from the pretrained Stable Video Diffusion model and encode the RGB image x, depth d, and surface normal n into the latent space using the frozen Stable Diffusion VAE. (b) Noise is sampled and added to the latent z. The image latent is replicated across the geometry slots (depth and normal), and further conditioned by concatenating the geometry latents. (c) We fine-tune only the GeoNeXt U-Net. (d) The model is optimized with the standard diffusion objective over image, depth, and normal latents to ensure fine-grained alignment and consistency between image and geometry.

Unified prediction

One model. Joint depth and normal prediction.

In the wild

Joint depth–normal prediction

Input
In-the-wild input

GeoWizard

Depth
GeoWizard depth
Normal
GeoWizard normal

GeoNeXt-WAN

Depth
GeoNeXt-WAN depth
Normal
GeoNeXt-WAN normal

Hover over any image to compare the same region across all predictions. Scroll to zoom.

Benchmark comparisons

Across scenes and modalities.

Each benchmark is evaluated against its corresponding ground truth. Depth and surface-normal results are presented separately to preserve the correct evaluation setting.

Depth benchmark

Hover over any image to compare the same region across all methods. Scroll to zoom.

Citation

BibTeX

@inproceedings{yang2026geonext,
  title     = {Video Generative Models as Geometry Learner},
  author    = {Yang, Haosen and Song, Jifei and Zhang, Zhensong and Zhu, Xiatian and Deng, Jiankang},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}
Enlarged comparison result