GeoNeXt

Video Generative Models as Geometry Learner

Homepage · Paper · Models · 🤗 Online Demo

Video Generative Models as Geometry Learner
Haosen Yang, Jifei Song, Zhensong Zhang, Xiatian Zhu*, Jiankang Deng*
ECCV 2026

This repository is the official implementation of "Video Generative Models as Geometry Learner".

Demo

The online demo runs on Hugging Face ZeroGPU. You may need to sign in to Hugging Face, and inference time depends on the current queue.

GeoNeXt_demo_web.mp4

Updates

• 2026/08/31: Inference code released.

Framework

Requirements

• Linux
• Python ≥ 3.10
• CUDA GPU with 24 GB VRAM (sufficient for inference)

Note: GeoNeXt-Wan and GeoNeXt-SVD should be installed in separate environments.

GeoNeXt-SVD

conda create -n geonext_svd python=3.10 -y
conda activate geonext_svd

git clone https://github.com/Creative-Intelligence-Studio/GeoNeXt.git
cd GeoNeXt
python -m pip install torch==2.3.1 torchvision==0.18.1 \
  --index-url https://download.pytorch.org/whl/cu121
python -m pip install -r requirements-svd.txt
python -m pip install -e .

GeoNeXt-Wan

GeoNeXt includes the tested DiffSynth inference implementation under diffsynth/; no separate DiffSynth checkout is required.

conda create -n geonext_wan python=3.10 -y
conda activate geonext_wan

git clone https://github.com/Creative-Intelligence-Studio/GeoNeXt.git
cd GeoNeXt
python -m pip install torch==2.3.1 torchvision==0.18.1 \
  --index-url https://download.pytorch.org/whl/cu121
python -m pip install -r requirements-wan.txt
python -m pip install -e .

The tested Wan dependency stack uses Diffusers 0.28.0, Transformers 4.40.1, PEFT 0.7.0, and Accelerate 0.29.3. Avoid upgrading these packages independently.

🤗 Pretrained Models

Our pretrained models are available on the Hugging Face Hub:

Version	Hugging Face Model	Depth	Normal	#Params
GeoNeXt-Wan	happy0612/GeoNeXt-Wan	✅	✅	1.42B
GeoNeXt-SVD	happy0612/GeoNeXt-SVD	✅	✅	1.52B

The selected GeoNeXt checkpoint, base model, and VAE are downloaded automatically on the first inference run and then reused from the Hugging Face cache.

To download both GeoNeXt checkpoints manually instead:

hf download happy0612/GeoNeXt --local-dir checkpoints

Local model paths can still be supplied with --checkpoint, --base-model, and --vae-model for offline use.

Inference

The --input argument accepts either one image or a directory. Supported image formats are JPEG, PNG, WebP, and BMP.

GeoNeXt-Wan

conda activate geonext_wan

python inference.py \
  --input assets/input \
  --output outputs/wan \
  --steps 5

GeoNeXt-SVD

conda activate geonext_svd

python inference.py \
  --backend svd \
  --input assets/input \
  --output outputs/svd \
  --steps 5

Run python inference.py --help for all shared options. Mesh export is optional and is enabled only when --export mesh is provided. Exported meshes use MoGe alignment by default. Use --align-space relative explicitly to export a mesh without MoGe.

Outputs

Both backends write the same directory structure:

outputs/<backend>/
├── depth_raw/         # Normalized disparity in [0, 1] (.npy)
├── depth_vis/         # Colorized depth maps (.png)
├── normal_raw/        # Surface normals in [-1, 1] (.npy)
├── normal_vis/        # RGB normal visualizations (.png)
└── geometry/<image>/  # Optional depth, intrinsics, and mesh files

Metric-scale alignment

Mesh export uses metric-space MoGe alignment by default. Activate the corresponding backend environment and install the pinned MoGe-2 integration once:

./scripts/setup_moge.sh third_party/MoGe

The setup script reuses the active environment's tested PyTorch installation and does not modify its PyTorch, CUDA, Transformers, Diffusers, or NCCL packages. Then export an aligned mesh with:

python inference.py \
  --backend wan \
  --input assets/input/case1.jpg \
  --output outputs/wan-moge \
  --export mesh

To export relative geometry without installing MoGe, add --align-space relative.

Web viewer

After exporting a mesh, start the viewer with:

python -m http.server 8000

Open http://localhost:8000/viewer/ and select the exported .ply file.

Citation

If GeoNeXt contributes to your work, please cite our paper:

@article{geonext2026,
  title   = {Video Generative Models as Geometry Learner},
  author  = {Yang, Haosen and Song, Jifei and Zhang, Zhensong and Zhu, Xiatian and Deng, Jiankang},
  journal = {arXiv preprint arXiv:2608.28549},
  year    = {2026}
}

Acknowledgements

This release builds on Wan, Stable Video Diffusion, Hugging Face Diffusers, DiffSynth, MoGe, and Trimesh. Please also follow the licenses and citation requirements of those projects. See NOTICE.md for third-party notices.