harry7557558/spirula-studio

Cross-vendor 3D Gaussian Splatting trainer - video to splat to mesh, Vulkan or CUDA.

What it solves

Spirula Studio provides a streamlined, all-in-one pipeline for 3D Gaussian Splatting. It removes the need for complex dependencies like Python, PyTorch, or separate Structure-from-Motion (SfM) tools like COLMAP, allowing users to transform raw photos or videos into 3D splats and textured meshes from a single binary.

How it works

The project uses a Vulkan-based compute backend to ensure cross-vendor GPU support (NVIDIA, AMD, Intel, and Apple Silicon). It implements a training strategy that combines MCMC, IGS+, and MRNF to produce sharper results with fewer artifacts. To maintain high efficiency, it uses quantized training, enabling the training of up to 10 million Gaussians within 8GB of VRAM. It also includes built-in tools for frame extraction, AI masking (via SAM), and native SfM for camera pose estimation.

Who it’s for

Artists, 3D developers, and researchers who want a high-performance 3D reconstruction pipeline that is easy to install and self-contained, regardless of their GPU hardware.

Highlights

  • Cross-Vendor GPU Support: Works on NVIDIA, AMD, Intel, and Apple GPUs via Vulkan.
  • VRAM Efficiency: Quantized training allows for large scenes (10M Gaussians) on consumer-grade hardware (8GB VRAM).
  • End-to-End Workflow: Integrated frame extraction, AI masking, and SfM, eliminating the need for external scripts.
  • Specialized Camera Support: Native support for 360° and equirectangular cameras without requiring undistortion.
  • Quality Enhancements: Uses modified Bilateral grid and PPISP for exposure and white balance correction.

Related

  • Project
    NVlabs/nvdiffrecmc

    A PyTorch implementation for jointly reconstructing 3D topology, materials, and lighting from multi-view images using Monte Carlo rendering and denoising.

  • Project
    apple-aiml-research/ml-hugs

    HUGS is a reference implementation for reconstructing an animatable human and their surrounding background scene from a single video using Gaussian Splatting.

  • Project
    ilastik/ilastik

    An interactive machine learning toolkit for segmenting, classifying, tracking, and counting cells and other experimental image data.

  • Project
    prs-eth/LitePT

    LitePT is a lightweight, high-performance 3D point cloud architecture that reduces parameter count and memory footprint while maintaining state-of-the-art accuracy for 3D vision tasks.