LoCoSplat: Real-Time Feed-Forward 3D Gaussian Splatting with Minimal 3D Reasoning

Sinan Wang*, Jinjin He*, Yuchen Sun, Duowen Chen, Shenyifan Lu, Bo Zhu (Georgia Institute of Technology; * equal contribution)

Paper (arXiv:2610.04351) · Code

LoCoSplat is a feed-forward 3D Gaussian Splatting model that replaces the learned 3D network of voxel-aligned methods with two fixed grids (voxel 0.001 and 0.128) read back at each anchor by a 0.14M-parameter pointwise MLP. With no learned 3D network and no dynamic sparse computation, the whole encoder runs as one fp16 CUDA graph: a 6-view scene is reconstructed in 33 ms on one RTX PRO 6000.

Checkpoints

All trained on RealEstate10K (256x256) for 150k steps, 4 GPUs x batch 1, with VolSplat's official recipe (config/experiment/locosplat_re10k.yaml in the code repository). Weights only (state_dict, optimizer states stripped), 397 MB each; 99.1 M parameters of which 86.6 M are the frozen Depth Anything V2 ViT-B backbone.

file	trained with	RE10K official index: PSNR / SSIM / LPIPS	sha256
locosplat_re10k_6v.ckpt	6 context views	31.51 / 0.942 / 0.074	03b03e4531167a65d8caa03dcf905212aaa4a2df408b3d4af6bd6d9adeb0485b
locosplat_re10k_12v.ckpt	12 context views	31.47 / 0.942 / 0.073	fffd4dc768fc047a95321078be2dd4b6580390dd728720dd0020438b8387ed76
locosplat_re10k_24v.ckpt	24 context views	30.57 / 0.934 / 0.082	790ce657dfe56a04a45a13fae4fee1528829eccef561b087d3cad2893a832e80

The 6-view checkpoint is also the model evaluated zero-shot on ACID (33.16 / 0.931 / 0.091) and the initialisation of the ScanNet fine-tuning in the paper.

Usage

git clone https://github.com/swang3081/LoCoSplat && cd LoCoSplat   # install: see README
hf download swang3081/LoCoSplat --local-dir models

scripts/eval_re10k.sh 6  models/locosplat_re10k_6v.ckpt          # plain fp32
FAST=1 scripts/eval_re10k.sh 6 models/locosplat_re10k_6v.ckpt   # deployed fp16 + CUDA-graph path
scripts/eval_re10k.sh 12 models/locosplat_re10k_12v.ckpt
scripts/eval_re10k.sh 24 models/locosplat_re10k_24v.ckpt

Each checkpoint file carries a locosplat dict with the training view count and the exact hydra overrides of its run.

Citation

@article{wang2026locosplat,
  title   = {LoCoSplat: Real-Time Feed-Forward 3D Gaussian Splatting with Minimal 3D Reasoning},
  author  = {Wang, Sinan and He, Jinjin and Sun, Yuchen and Chen, Duowen and Lu, Shenyifan and Zhu, Bo},
  journal = {arXiv preprint arXiv:2610.04351},
  year    = {2026}
}

Acknowledgements

Built on VolSplat, DepthSplat, UniMatch and Depth Anything V2.