EcoSplat: Efficiency-controllable Feed-forward 3D Gaussian Splatting from Multi-view Images

CVPR 2026 (Highlight)

Paper · Project Page · Code

EcoSplat is a feed-forward 3D Gaussian Splatting model whose compute/quality trade-off is controllable at inference time via a single primitive budget, the protect rate κ (lower κ → fewer rendered Gaussians). This repository hosts the efficiency-controllable (IGF) checkpoints.

Checkpoints

File	Base	Train data
ecosplat-spfsplat-re10k.ckpt	SPFSplat	RealEstate10K
ecosplat-zpressor-re10k.ckpt	ZPressor / MVSplat	RealEstate10K

License & provenance

Released under CC BY-NC 4.0 (non-commercial; see LICENSE). © 2026 Minh-Quan Viet Bui.

• ecosplat-spfsplat-re10k.ckpt is fine-tuned from MASt3R (Naver), which is non-commercial (CC BY-NC-SA 4.0) — that restriction applies to this checkpoint.
• ecosplat-zpressor-re10k.ckpt is fine-tuned from MVSplat + ZPressor (MIT).

Usage

Download a checkpoint:

from huggingface_hub import hf_hub_download
ckpt = hf_hub_download("quan5609/EcoSplat", "ecosplat-spfsplat-re10k.ckpt")

or

wget https://huggingface.co/quan5609/EcoSplat/resolve/main/ecosplat-spfsplat-re10k.ckpt

Then follow the inference / evaluation instructions for the matching base, on the corresponding code branch:

• SPFSplat base — eco_spfsplat
• ZPressor / MVSplat base — eco_zpressor

Set the primitive budget at test time with +model.encoder.igf.inference_rho=<κ> (e.g. sweep 0.7 → 0.1 → 0.02).

Citation

@article{park2025ecosplat,
  title={EcoSplat: Efficiency-controllable Feed-forward 3D Gaussian Splatting from Multi-view Images},
  author={Park, Jongmin and Bui, Minh-Quan Viet and Bello, Juan Luis Gonzalez and Moon, Jaeho and Oh, Jihyong and Kim, Munchurl},
  journal={arXiv preprint arXiv:2512.18692},
  year={2025}
}

Acknowledgements

Built on SPFSplat, ZPressor, NoPoSplat, pixelSplat, MVSplat, DepthSplat, DUSt3R, and CroCo.