[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1qc40bqfemmz9":3},{"code":4,"message":5,"data":6,"timestamp":46},10000,"success",{"full_name":7,"position":8,"description":9,"language":10,"stars":11,"forks":12,"stars_gained":13,"age_days":14,"hn_threads":15,"one_liner":22,"summary_md":23,"topic":24,"github_topics":27,"license":42,"homepage":43,"updated_at":44,"history":45},"harry7557558\u002Fspirula-studio",4,"Cross-vendor 3D Gaussian Splatting trainer - video to splat to mesh, Vulkan or CUDA.","C++",1862,170,1184,903,[16],{"id":17,"title":18,"points":19,"comments":20,"posted_at":21},49895287,"Spirula-studio: 3D Gaussian Splatting trainer video to splat to mesh",5,0,"2026-09-29","Trains 3D Gaussian Splatting models from photos or video inside one binary, without Python, PyTorch or COLMAP.","## The problem\nGetting a 3D Gaussian Splatting model out of raw footage normally means assembling a stack: COLMAP for structure-from-motion, a PyTorch reference trainer, ffmpeg for frame extraction, maybe a separate masking model, and then something else for meshing. Each piece has its own GPU assumptions, and most of them assume an NVIDIA card. Spirula Studio bundles the whole path — video to splat to textured mesh — into a single native binary with a GUI and a CLI.\n\n## What's different\nThe usual reference pipeline is CUDA-only and Python-first; Spark, gsplat and the original INRIA code all live in that world. Spirula's main differentiator is the Vulkan compute backend, which runs the same training on NVIDIA, AMD, Intel and Apple Silicon, with the CUDA path kept as a legacy option. The second differentiator is memory: quantized training is claimed to hold 10M full-SH Gaussians in 8 GB of VRAM, which matters for large scenes on consumer cards. Third, fisheye and equirectangular 360° input is handled natively rather than requiring prior undistortion. The built-in SfM, AI masking and frame extraction remove the COLMAP and ffmpeg steps, but those are convenience wins rather than architectural ones.\n\n## Who it's for\nPhotogrammetry and 3D-capture practitioners without NVIDIA hardware — AMD, Intel or Apple Silicon users who currently cannot run the reference trainer at all — and people who need 360° or fisheye captures without a preprocessing detour. It also suits anyone doing cloud or headless training, since the CLI can serve a viewer over an SSH tunnel. Look elsewhere if you need to modify the training loop: this is a finished application, not a research framework with a Python API. Teams already deep in the PyTorch ecosystem with NVIDIA cards have less reason to switch.\n\n## Cost of adoption\nThe GUI downloads are the low-friction path; building requires the Vulkan SDK (MoltenVK is fetched automatically on macOS) plus CMake\u002FNinja, or a recent CUDA toolkit for the legacy backend. The `-DSS_ENABLE_PATENTED=ON` flag enables GPU video decoding but puts AVC\u002FHEVC patent compliance on you. Masking pulls a SAM checkpoint from Meta on first use, and SAM 3 is under a non-standard license, so check that before commercial use. The whole project is GPL-3.0. It is effectively a one-person project, and the README says so directly — expect occasionally slow issue responses.\n\n## Verdict\nSpirula Studio is not new: it has been public since May 2024, and its recent momentum was a monthly C++ board appearance at #4 with +1,184 stars rather than a sudden viral spike. That pattern suggests a project that has been steadily accumulating users and features — the changelog shows work on metric scale, LoMa features and macOS in recent months. The cross-vendor Vulkan bet is the real reason to care, and it is the one thing the CUDA-centric incumbents do not offer. The risks are honest ones: single maintainer, GPL, and optional components with their own licensing. If you are on non-NVIDIA hardware or need 360° input, it is worth a download; if you need to fork the trainer, it is not.",{"slug":25,"name":26},"machine-learning","Machine Learning",[28,29,30,31,32,33,34,35,36,37,38,39,40,41],"360-camera","3d-gaussian-splatting","3d-reconstruction","3dgs","3dgs-acceleration","computer-vision","cross-vendor","cuda","gaussian-splatting","mesh-generation","quantized-training","structure-from-motion","vram-optimization","vulkan","GPL-3.0",null,"2026-10-24T18:37:28Z",[],1792883104]