InfiniSplat: Implicit Gaussian Decoding for Large-Baseline Monocular View Synthesis

Project Page | Code | Gallery

Overview

InfiniSplat reconstructs a 3D Gaussian scene representation from a single image. It supports RGB-only reconstruction and depth-sensor-guided reconstruction from a spatially aligned RGB-depth pair.

Checkpoint	Input	Output
infinisplat_rgb.ckpt	RGB image	3D Gaussian Splatting scene
infinisplat_lidar.ckpt	RGB image + aligned depth	3D Gaussian Splatting scene

The released checkpoints contain the complete inference weights required by their respective encoders.

Usage

Clone the InfiniSplat repository, follow the installation guide, and download the checkpoints:

bash scripts/download_checkpoints.sh

RGB-only inference:

python -m src.demo.infer_batch_images --input examples/data/rgb_demo/pexels-masi.jpg

Depth-sensor-guided inference:

python -m src.demo.infer_batch_images \
  --mode lidar \
  --input examples/data/lidar_demo/eth3d_kicker.png

The input path may also be a directory for batch inference. In LiDAR mode, RGB and depth files are paired by filename stem. See the full inference guide for supported depth formats, camera intrinsics, outputs, and optional arguments.

Outputs

Inference always exports a Gaussian PLY. Novel-view MP4 rendering is available when gsplat is installed, and interactive HTML export is available when the PlayCanvas splat-transform CLI is installed.

Acknowledgments

InfiniSplat builds on DINOv3, Depth Pro, InfiniDepth, and gsplat. We thank their authors for their excellent work.