> GitHub Trending deep dive | 2026-09-23 > Project: harry7557558/spirula-studio > Language: C++ | License: GPLv3 | Stars today: +99
TL;DR
Spirula Studio is not yet another 3DGS trainer — it is a de-Pythonized rebuild of the entire 3D reconstruction toolchain. While everyone else cycles through PyTorch + COLMAP + endless pip installs, the author compiled the complete video-to-mesh pipeline into one executable — and it supports NVIDIA, AMD, Intel, and Apple GPUs.
The Problem: A Typical 3DGS Workflow Breakdown
Anyone who has done 3D Gaussian Splatting knows the drill:
1. Install CUDA, PyTorch, and various Python dependencies (two hours) 2. Install COLMAP for SfM (another hour, often fails to build) 3. Write scripts for frame extraction and masking 4. Train (OOM, tweak batch size) 5. Export mesh (install yet another tool)
Then switch machines and repeat everything. Spirula Studio's author, harry7557558, clearly got fed up. His solution was blunt: compile everything into a single binary. Download, unzip, double-click, done.
Core Innovation: Engineering, Not Algorithms
1. Cross-Vendor Vulkan Backend
This is Spirula's most unusual trait. 3DGS training is almost entirely CUDA-monopolized — mainstream implementations (original 3DGS, gsplat, INRIA's reference code) assume an NVIDIA card.
Spirula uses Vulkan compute as its training backend, unlocking:
| GPU vendor | Supported | |------------|-----------| | NVIDIA | ✅ | | AMD | ✅ | | Intel | ✅ | | Apple Silicon | ✅ (via MoltenVK) |
This isn't "theoretical support" — the Vulkan backend is the recommended option, more mature than the CUDA backend. The README states: "Vulkan backend is the most tested."
Analogy: it's like a video encoder that skips NVIDIA NVENC and uses Vulkan Compute so every GPU can hardware-encode. Not impossible — just something nobody bothered to do, because the CUDA ecosystem is too comfortable.
2. 10 Million Gaussians in 8GB VRAM
3DGS VRAM usage is notoriously high: 10 million full-SH3 Gaussians need at least 24GB in the original implementation. Spirula squeezes the same scale into 8GB via quantized training. Not magic, just engineering:
- Quantized training: low-precision storage during gradient accumulation
- Adaptive strategy: dynamic batch sizing based on VRAM
- MCMC/IGS+/MRNF combined: fewer floats, sharper results
- SfM (structure from motion): no COLMAP needed
- AI masking: automatic foreground/background separation
- Video frame extraction: feed video directly
- 360° camera support: native fisheye and equirectangular projection
- Mesh export: mesh out right after training
- Metric scale recovery: new Sept 10 feature, restores real-world scale from video metadata
- 3D photography hobbyists: AMD/Intel GPU owners finally don't need to borrow an NVIDIA machine
- Surveying/GIS professionals: 360° cameras + metric scale recovery for usable scene reconstructions
- E-commerce / artifact scanning: video in, mesh out, zero Python knowledge
- 3DGS researchers: a binary is more reliable than pip install for quick baseline reproduction
- Those who need to hack on training algorithms (source is buildable, but C++ is a higher bar than Python)
- Those chasing peak performance (Vulkan is a few percent slower than CUDA)
- Anyone with a working CUDA workflow (no reason to switch)
- Project: github.com/harry7557558/spirula-studio
- Online demo: spirula.studio
- Web viewer: harry7557558.github.io/spirula-studio/viewer
- Original 3DGS paper: Kerbl et al., "3D Gaussian Splatting for Real-Time Radiance Field Rendering", SIGGRAPH 2023
Analogy: like smartphone HDR+ — not a better sensor, but computational photography squeezing every photon.
3. Fully Integrated Pipeline
Spirula builds in:
Analogy: from "building your own PC" to "buying a prebuilt." No single part is the strongest, but integration is the best.
Why Vulkan?
3DGS training is dominated by large matrix multiplies and Gaussian rasterization, where CUDA has an absolute edge — cuDNN, CUTLASS, and warp-level primitives are NVIDIA's moat. But Spirula chose Vulkan compute for pragmatic reasons:
1. Coverage: NVIDIA holds only ~20-30% of PC GPUs. The other 70%+ are AMD/Intel/Apple. CUDA-only 3DGS stays a niche tool forever. 2. Deployment: Vulkan is cross-platform (Windows/Linux/macOS). CUDA works best on Linux, tolerably on Windows, not at all on macOS. 3. The future: Apple Silicon GPU compute is now non-trivial. An M3 Max can train mid-size 3DGS, but CUDA blocks that path entirely.
The cost: Vulkan compute's ecosystem is far less mature. No cuDNN-grade libraries mean many operations require hand-written shaders, and Spirula's codebase is much larger than a pure CUDA implementation would be.
A classic trade of short-term cost for long-term coverage.
Timeline: From CUDA-Only to All Vendors
| Date | Feature | |------|---------| | 2026-07-22 | Vulkan backend launched, NVIDIA/AMD/Intel support | | 2026-08-08 | End-to-end workflow (frame extraction, AI masking, SfM, meshing) | | 2026-08-08 | Multi-language UI (Simplified/Traditional Chinese, Japanese, Korean, major European languages) | | 2026-08-14 | macOS / Apple Silicon verified | | 2026-09-03 | LoMa feature support (SfM on hard datasets) | | 2026-09-10 | Metric scale recovery from video metadata |
From "CUDA only" to "all vendors + full pipeline" in two months — a blistering pace.
Comparison with Mainstream 3DGS Tools
| Tool | Backend | Pipeline | Deployment | 360° | Platforms | |------|---------|----------|------------|------|-----------| | Original 3DGS (INRIA) | CUDA | Training only | Python scripts | ❌ | Linux mainly | | gsplat | CUDA | Training only | Python package | ❌ | Linux/Win | | Luma AI | Cloud | Full | Closed-source SaaS | ✅ | Web | | Postshot | CUDA | Full | Closed-source beta | ❌ | Win | | Spirula | Vulkan/CUDA | Full | Open-source binary | ✅ | Win/Linux/macOS |
Spirula's positioning is clear: open source, full pipeline, all vendors. In that quadrant, it has almost no competitors.
Who Should Use It?
Not for:
My Take
Spirula Studio embodies an underrated engineering philosophy: the best tool isn't the most powerful one — it's the one with the highest coverage.
The 3DGS field doesn't lack CUDA implementations or paper reproductions; it lacks something that "just works on double-click." Spirula did the thankless job of rewriting the entire training stack in Vulkan so the 70% of non-NVIDIA users can play with 3DGS too.
That decision isn't popular in academia (Vulkan compute doesn't make papers) or in startups (SaaS monetizes faster). But it solves a real problem: 3D reconstruction tools have too high a barrier.
The author, harry7557558, is an independent developer with no big-company backing. One person rewriting an entire toolchain in C++ has the same spirit as HandBrake wrapping ffmpeg into a cross-platform GUI back in the day — not sexy, but important.