WorldRoPE

Probabilistic World-Centric Rotary Embedding for Multi-View Attention

Mijeong Kim · Kyujin Lim · Bohyung Han
Seoul National University

arXiv · Paper · Project Page · Code

WorldRoPE represents image tokens as probabilistic 3D Gaussian supports and integrates rotary embeddings over their projection into a query view. This repository provides LVSM novel view synthesis checkpoints for WorldRoPE, along with the RayRoPE and URoPE baselines.

Checkpoints

Method	Dataset	Weights	Configuration
RayRoPE (ECCV 2026, spotlight)	RE10K	Download	config.yaml
URoPE (ECCV 2026, spotlight)	RE10K	Download	config.yaml
WorldRoPE	RE10K	Download	config.yaml
WorldRoPE	Objaverse	Download	config.yaml

hf download mijeongkim/WorldRoPE --local-dir checkpoints

Each checkpoint includes the model tensors and training step, paired with its matching configuration. All four checkpoints are from step 80,000. WorldRoPE and URoPE use 24.5M parameters; RayRoPE uses the official 47.4M small pretrained model from RayRoPE.

Evaluation

Install the code release and prepare the dataset as described in its README. Then verify and evaluate:

python scripts/check_checkpoint.py \
  --checkpoint checkpoints/worldrope_lvsm/re10k/model.pt \
  --config checkpoints/worldrope_lvsm/re10k/config.yaml

bash scripts/nvs.sh --dataset re10k --method worldrope --gpus 0 --test \
  --model-config checkpoints/worldrope_lvsm/re10k/config.yaml \
  --data-root /path/to/re10k \
  --checkpoint checkpoints/worldrope_lvsm/re10k/model.pt \
  --output-dir /path/to/new-evaluation-output

For RayRoPE or URoPE, select --method rayrope or --method urope and use the corresponding checkpoint and configuration from the table. For Objaverse, use --dataset objaverse with its checkpoint/configuration and --fast-dir /path/to/objaverse-packed instead of --data-root.

Results

Reported WorldRoPE results:

Dataset	PSNR ↑	SSIM ↑	LPIPS ↓
RE10K	26.75	0.846	0.073
Objaverse	22.52	0.896	0.113

Citation

@article{kim2026worldrope,
  title={WorldRoPE: Probabilistic World-Centric Rotary Embedding for Multi-View Attention},
  author={Kim, Mijeong and Lim, Kyujin and Han, Bohyung},
  journal={arXiv preprint arXiv:2610.16337},
  year={2026}
}

Acknowledgements

The baselines build on RayRoPE and URoPE. See LICENSE and upstream notices; RayRoPE retains its MIT license notice.