globalsplat

GlobalSplat — pretrained weights

Feed-forward, generalizable 3D Gaussian Splatting via global scene tokens (align first, decode later). These are the released RealEstate10K checkpoints for the GlobalSplat paper.

• Project page: https://r-itk.github.io/globalsplat/
• Paper: https://arxiv.org/abs/2604.15284

Checkpoints

The -noopacity 16K (trained with the released code, decoder opacity regularizer disabled) outperforms the paper default at every context count: 12v 29.10 / 24v 28.95 / 36v 28.76 PSNR.

All files are weights-only checkpoints (state_dict + Lightning version; the frozen VGG perceptual net and optimizer state are stripped).

Usage

With the GlobalSplat code:

# 16K paper default
python -m globalsplat.main +experiment=re10k_16k dataset=re10k_eval_ctx24 \
    mode=test checkpointing.load=globalsplat-re10k-16k.ckpt
# 2K (sets M_max=2) and 32K
python -m globalsplat.main +experiment=re10k_2k  ... checkpointing.load=globalsplat-re10k-2k.ckpt
python -m globalsplat.main +experiment=re10k_32k ... checkpointing.load=globalsplat-re10k-32k.ckpt

License

PolyForm Noncommercial 1.0.0 with additional terms (see LICENSE and ADDITIONAL_TERMS.md): noncommercial research / academic / personal use only; commercial use requires a separate written license; military and surveillance use are prohibited. Source-available, not an OSI open-source license.

Citation

@article{itkin2026globalsplat,
  title={GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens},
  author={Itkin, Roni and Issachar, Noam and Keypur, Yehonatan and Chen, Xingyu and Chen, Anpei and Benaim, Sagie},
  journal={arXiv preprint arXiv:2604.15284},
  year={2026}
}