spirula-studio

Product Introduction

1. Definition: Spirula Studio is a cross-platform, cross-vendor 3D Gaussian Splatting (3DGS) training and reconstruction software. It is a self-contained binary application that performs the entire pipeline from 2D video or image input to a trained 3D Gaussian Splatting model and textured mesh output.
2. Core Value Proposition: It exists to democratize high-quality 3D asset creation by eliminating complex, fragmented software dependencies. Its primary value is providing a single, efficient tool for 3D Gaussian Splatting training that works on any major GPU (NVIDIA, AMD, Intel, Apple) without requiring separate installations of Python, PyTorch, or COLMAP.

Main Features

1. Cross-Vendor Vulkan Compute Backend: The software utilizes a custom Vulkan compute backend for all core training operations. This enables hardware-agnostic acceleration, allowing it to run on NVIDIA, AMD, Intel Arc, and Apple Silicon GPUs with near-identical performance and feature sets. It works via the standard Vulkan API on Windows/Linux and MoltenVK on macOS.
2. Extreme VRAM Efficiency with Quantized Training: Spirula Studio employs a quantized training strategy that drastically reduces GPU memory consumption. This allows it to train models with up to 10 million Spherical Harmonics (SH) degree 3 Gaussians within just 8 GB of VRAM, enabling high-detail reconstruction on consumer-grade graphics cards.
3. Integrated End-to-End Pipeline: The tool features a built-in, lightning-fast Structure-from-Motion (SfM) module for camera pose estimation, AI-powered masking (using SAM 2.1/3), and GPU-accelerated video frame extraction. This creates a complete, dependency-free workflow from raw video to final 3D asset.
4. Native 360° and Fisheye Camera Support: Unlike many 3D reconstruction tools that require pre-undistorted images, Spirula Studio has native support for equirectangular 360° imagery and fisheye camera models. Users can directly load footage from 360 cameras without manual preprocessing.
5. Advanced Training Strategy (MCMC/IGS+/MRNF Hybrid): It implements a proprietary training strategy that combines principles from MCMC (Markov Chain Monte Carlo) densification, IGS+ (Instant Gaussian Splatting), and MRNF (Multi-Resolution Neural Feature) fields. This results in sharper reconstructions with fewer visual artifacts like floaters, effective for both object-scale and large-scale scenes.
6. Automatic Color Correction and Meshing: The pipeline includes a modified bilateral grid and PPISP (Per-Pixel Illumination and Shading Processing) module for automatic exposure and white balance correction. It also features built-in tools for generating depth/normal maps, converting splats to textured meshes, and creating skyboxes.

Problems Solved

1. Pain Point: The complexity and fragmentation of the standard 3D Gaussian Splatting pipeline, which typically requires managing multiple separate tools (FFmpeg, COLMAP, PyTorch with CUDA, meshing tools) and deep technical expertise to install and configure.
2. Target Audience: Independent 3D creators, game developers, VR/AR content producers, architectural visualization artists, and researchers who need to create 3D assets from real-world captures but lack access to high-end NVIDIA hardware or the time to manage complex technical stacks.
3. Use Cases: Creating 3D assets for games and VR from smartphone video; generating digital twins of rooms or objects for e-commerce; rapid prototyping for product design using video reference; academic research in 3D computer vision without CUDA-dependent infrastructure; creating immersive 360° environment maps for virtual tours.

Unique Advantages

1. Differentiation: Compared to traditional PyTorch-based 3DGS implementations (like original 3D Gaussian Splatting or Instant-NGP), Spirula Studio is a standalone binary with no Python dependencies. Versus cloud-based services, it offers full privacy, offline operation, and no per-model fees. Its cross-vendor support is unique in the space, breaking the NVIDIA CUDA monopoly.
2. Key Innovation: The development of a fully-featured, production-grade 3D Gaussian Splatting training engine on Vulkan Compute is its core technical innovation. This, combined with its quantized training algorithm and integrated SfM/AI masking, creates an unprecedented all-in-one, hardware-accessible solution for 3D reconstruction.

Frequently Asked Questions (FAQ)

1. What GPUs does Spirula Studio support? Spirula Studio supports NVIDIA GPUs (via CUDA or Vulkan), AMD GPUs (via Vulkan), Intel Arc GPUs (via Vulkan), and Apple Silicon Macs (via MoltenVK). It is one of the only 3D Gaussian Splatting trainers with true cross-vendor compatibility.
2. Do I need to install COLMAP or PyTorch to use Spirula Studio? No. Spirula Studio is a completely self-contained application. It has its own built-in, fast SfM (Structure-from-Motion) system for estimating camera poses and does not require Python, PyTorch, COLMAP, or FFmpeg to be installed on your system for core functionality.
3. How much VRAM is needed to train a model with Spirula Studio? Thanks to its quantized training techniques, Spirula Studio is highly VRAM efficient. It can train a detailed model with 10 million Gaussians using approximately 8 GB of GPU memory. Less complex scenes will require significantly less VRAM.
4. Can Spirula Studio process 360-degree video? Yes. A key feature of Spirula Studio is native support for equirectangular 360° imagery and fisheye lens models. You can directly input video or images from a 360 camera without the need for prior undistortion or conversion.
5. What is the difference between the Vulkan and CUDA backends? The Vulkan backend is cross-platform and works on all major GPU brands. The CUDA backend is legacy and only for NVIDIA GPUs. Both offer similar training performance. The Vulkan backend is recommended as it includes the latest features like integrated SfM and video decoding, and is the only option for macOS and non-NVIDIA GPUs.