AirSplat: Alignment and Rating for Robust Feed-Forward 3D Gaussian Splatting

ECCV 2026 · Paper · Project page · Code

Minh-Quan Viet Bui*, Jaeho Moon*, Munchurl Kim — KAIST (*equal contribution)

AirSplat adapts the Depth Anything 3 foundation model into a robust feed-forward 3D Gaussian Splatting model for novel-view synthesis, using Self-Consistent Pose Alignment (training-time feedback against pose–geometry misalignment) and Rating-based Opacity Matching (teacher-guided filtering of inconsistent primitives).

Files

file	size	contents
airsplat-dl3dv.ckpt	5.6 GB (fp32)	full DA3-Giant backbone + AirSplat Gaussian heads, PyTorch state_dict (no optimizer state)

The checkpoint is self-contained; the code never downloads the backbone separately.

Usage

git clone https://github.com/KAIST-VICLab/airsplat && cd airsplat
pip install -r requirements.txt   # see README for the PyTorch/CUDA step
wget -P pretrained_weights https://huggingface.co/quan5609/AirSplat/resolve/main/airsplat-dl3dv.ckpt
CKPT=pretrained_weights/airsplat-dl3dv.ckpt ./eval_dl3dv_24v.sh

License

CC BY-NC 4.0 — non-commercial use only. These weights are a derivative of the DA3-Giant weights (Copyright 2025 The Depth Anything 3 Team, CC BY-NC 4.0) and embed that backbone. The AirSplat source code is MIT-licensed separately.

Citation

@inproceedings{bui2026airsplatalignmentratingrobust,
    title={AirSplat: Alignment and Rating for Robust Feed-Forward 3D Gaussian Splatting},
    author={Minh-Quan Viet Bui and Jaeho Moon and Munchurl Kim},
    booktitle={ECCV},
    year={2026}
}