OmniCam — Pretrained Checkpoints

Pretrained weights for OmniCam: Omni-Camera Trajectory Generation via Geometry-Grounded Pose Token Learning.

Given a single 360° panorama and a textual trajectory description, OmniCam autoregressively generates a spatially aware SE(3) camera-pose sequence, supporting four camera behaviors — Target, Surround, Reconstruct, Wander.

• 💻 Code, dataset pipeline & evaluation benchmark: https://github.com/ZhenyangLiu/OmniCam
• 🌐 Project page: https://zhenyangliu.github.io/OmniCam/

Files

File	Selected by	Size
best_ate.safetensors	best trajectory accuracy (ATE) on held-out scenes	5.1 GB
best_loss.safetensors	lowest validation loss	5.1 GB

We recommend best_ate.safetensors for trajectory-quality evaluation.

Usage

pip install huggingface_hub
huggingface-cli download ZhenyangLiu/OmniCam best_ate.safetensors --local-dir checkpoints

# in the OmniCam repo
cd omnicam
RESUME_PATH=../checkpoints/best_ate.safetensors bash scripts/eval.sh

Citation

@inproceedings{omnicam2026,
  title     = {OmniCam: Omni-Camera Trajectory Generation via
               Geometry-Grounded Pose Token Learning},
  author    = {Anonymous Authors},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}