[NeurIPS 2025 Spotlight]
Orient Anything V2: Unifying Orientation and Rotation Understanding

Zehan Wang^1* · Ziang Zhang^1* · Jiayang Xu¹ · Jialei Wang¹ · Tianyu Pang² · Chao Du² · Hengshuang Zhao³ · Zhou Zhao¹

¹Zhejiang University ²SEA AI Lab ³HKU

*Equal Contribution

Orient Anything V2, a unified spatial vision model for understanding orientation, symmetry, and relative rotation, achieves SOTA performance across 14 datasets.

News

• 2025-12-12: 🔥Paper, Project Page, Code, Training Data, Model Checkpoint, and Demo have been released!

• 2025-09-18: 🔥Orient Anything V2 has been accepted as a Spotlight @ NeurIPS 2025!

Pre-trained Model Weights

We provide pre-trained model weights and are continuously iterating on them to support more inference scenarios:

Model	Params	Checkpoint
Orient-Anything-V2	5.05 GB	Download

Quick Start

1 Dependency Installation

conda create -n orianyv2 python=3.11
conda activate orianyv2
pip install -r requirements.txt

2 Gradio App

Start gradio by executing the following script:

python app.py

then open GUI page(default is https://127.0.0.1:7860) in web browser.

or, you can try it in our Huggingface-Space

3 Python Scripts

importnumpyasnpfromPILimportImageimporttorchimporttempfileimportosfrompathsimport*fromvision_towerimportVGGT_OriAny_Reffrominferenceimport*fromapp_utilsimport*mark_dtype=torch.bfloat16iftorch.cuda.get_device_capability()[0] >=8elsetorch.float16# device = 'cuda:0'device=torch.device('cuda'iftorch.cuda.is_available() else'cpu')
ifos.path.exists(LOCAL_CKPT_PATH):
ckpt_path=LOCAL_CKPT_PATHelse:
fromhuggingface_hubimporthf_hub_downloadckpt_path=hf_hub_download(repo_id="Viglong/Orient-Anything-V2", filename=HF_CKPT_PATH, repo_type="model", cache_dir='./', resume_download=True)
model=VGGT_OriAny_Ref(out_dim=900, dtype=mark_dtype, nopretrain=True)
model.load_state_dict(torch.load(ckpt_path, map_location='cpu'))
model.eval()
model=model.to(device)
print('Model loaded.')
@torch.no_grad()defrun_inference(pil_ref, pil_tgt=None, do_rm_bkg=True):
ifpil_tgtisnotNone:
ifdo_rm_bkg:
pil_ref=background_preprocess(pil_ref, True)
pil_tgt=background_preprocess(pil_tgt, True)
else:
ifdo_rm_bkg:
pil_ref=background_preprocess(pil_ref, True)
try:
ans_dict=inf_single_case(model, pil_ref, pil_tgt)
exceptExceptionase:
print("Inference error:", e)
raisegr.Error(f"Inference failed: {str(e)}")
defsafe_float(val, default=0.0):
try:
returnfloat(val)
except:
returnfloat(default)
az=safe_float(ans_dict.get('ref_az_pred', 0))
el=safe_float(ans_dict.get('ref_el_pred', 0))
ro=safe_float(ans_dict.get('ref_ro_pred', 0))
alpha=int(ans_dict.get('ref_alpha_pred', 1))
ifpil_tgtisnotNone:
rel_az=safe_float(ans_dict.get('rel_az_pred', 0))
rel_el=safe_float(ans_dict.get('rel_el_pred', 0))
rel_ro=safe_float(ans_dict.get('rel_ro_pred', 0))
print("Relative Pose: Azi",rel_az,"Ele",rel_el,"Rot",rel_ro)
image_ref_path='assets/examples/F35-0.jpg'image_tgt_path='assets/examples/F35-1.jpg'# optionalimage_ref=Image.open(image_ref_path).convert('RGB')
image_tgt=Image.open(image_tgt_path).convert('RGB')
run_inference(image_ref, image_tgt, True)

Evaluate Orient-Anything-V2

Data Preparation

Download the absolute orientation, relative rotation, and symm-orientation test datasets from Huggingface Dataset.

# set mirror endpoint to accelerate# export HF_ENDPOINT='https://hf-mirror.com'
huggingface-cli download --repo-type dataset Viglong/OriAnyV2_Inference --local-dir OriAnyV2_Inference

Use the following command to extract the dataset:

cd OriAnyV2_Inference
forfin*.tar.gz;do
tar -xzf "$f"done

Modify DATA_ROOT in paths.py to point to the dataset root directory(/path/to/OriAnyV2_Inference).

Evaluate with torch-lightning

To evaluate on test datasets, run the following code:

python eval_on_dataset.py

Train Orient-Anything-V2

We use FLUX.1-dev and Hunyuan3D-2.0 to generate our training data and render it with Blender. We provide the fully rendered data, which you can obtain from the link below.

Assets	Disk Space	Download Link
Images and 3D assets in the data pipeline	2 TB	Hunyuan3D-FLUX-Gen
Final Rendering Data	25 GB	Training Dataset

To store all this data, we recommend having at least 2TB of free disk space on your server.

We are currently organizing the complete data construction pipeline and training code for Orient-Anything-V2 — stay tuned.

Acknowledgement

We would like to express our sincere gratitude to the following excellent works:

• VGGT
• FLUX
• Hunyuan3D-2.0
• Blender
• rembg

Citation

If you find this project useful, please consider citing:

@inproceedings{wangorient,
title={Orient Anything V2: Unifying Orientation and Rotation Understanding},
author={Wang, Zehan and Zhang, Ziang and Xu, Jiayang and Wang, Jialei and Pang, Tianyu and Du, Chao and Zhao, Hengshuang and Zhao, Zhou},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems}
}