Open Source
Explore the latest AI open-source projects from GitHub and HuggingFace.
Explore the latest AI open-source projects from GitHub and HuggingFace.
Spirula Studio, formerly named spirulae-splat, is a GPL-3.0 licensed application from GitHub developer harry7557558 that trains 3D Gaussian Splatting scenes end to end, turning raw photos or video into a splat and then a textured mesh from one self-contained binary, without a separate Python or PyTorch environment or a standalone COLMAP install. It runs cross-vendor through Vulkan compute on NVIDIA, AMD, Intel, and Apple Silicon GPUs, and also offers a legacy CUDA backend for NVIDIA hardware; the two backends produce the same training and meshing results, with speed differing by only a few percent depending on the GPU driver. A single training strategy combines the advantages of the MCMC, IGS+, and MRNF Gaussian Splatting variants, and quantized training is described as fitting up to 10 million SH3 Gaussians inside 8 GB of VRAM. The application has native support for 360-degree and equirectangular camera inputs without a separate undistortion step, along with a modified bilateral grid and PPISP method for exposure and white-balance correction. It bundles its own structure-from-motion solver, AI-based masking, and video frame extraction, so a dataset can go from raw capture to training without external tools; the AI masking downloads a Segment Anything Model checkpoint on first use, either SAM 2.1 under Apache-2.0 or SAM 3 under Meta's own non-standard license, and the checkpoints are never bundled with the app. An optional build flag turns on GPU-accelerated video decoding, cutting frame extraction time by about 15 times versus shelling out to ffmpeg, though this path uses AVC/HEVC hardware decoding that carries its own third-party patent exposure the builder has to manage. Prebuilt binaries cover Windows, Linux, and macOS, including Apple Silicon, and a CLI mode can train on a remote or cloud GPU while serving a live progress viewer over an SSH-forwarded HTTP port. The interface and CLI support thirteen languages, and recent updates have added telemetry-based metric-scale recovery for datasets, LoMa feature matching for harder structure-from-motion cases, and editing tools for splats, meshes, and masks with render export.
graphdeco-inria
Original reference implementation of 3D Gaussian Splatting for real-time radiance field rendering
ahujasid
Connect Blender to Claude AI via MCP for natural-language-driven 3D scene creation and manipulation.
Tencent Hunyuan
Tencent's open-source diffusion system that turns a single image or text prompt into high-resolution, textured 3D assets.