One app covers the path from footage to a textured mesh
Spirula Studio takes raw photos or video, solves the camera positions, trains a 3D Gaussian splat, and can turn the result into a textured mesh. That is the reason to care. A typical open 3DGS workflow makes you connect several programs and scripts before you can inspect the scene. Here, frame extraction, masking, structure from motion, training, viewing, and meshing sit in one GUI, with a CLI for remote jobs.
The README says the Vulkan backend runs on NVIDIA, AMD, Intel, and Apple GPUs. It also claims support for fisheye and 360-degree input without an undistortion pass, plus training of as many as 10 million full-SH Gaussians in 8 GB of VRAM. Those are project claims rather than results from our container. Still, they explain the product's appeal: Spirula Studio is trying to make capture-to-splat work feel like an application, not a research environment.
Vulkan is the main route, while CUDA is the legacy choice
The documentation recommends Vulkan and describes CUDA as the older NVIDIA-only backend. Both expose training and meshing, but Vulkan also carries the built-in structure-from-motion, video extraction, and AI masking workflow. Windows, Linux, and macOS builds are published on the releases page. On a remote GPU, the CLI can serve a browser viewer that you reach through an SSH port forward.
Source builders have more choices to make. CMake and Ninja sit underneath the Vulkan route, while CUDA needs a compatible toolkit and compiler. Turning on GPU video decoding avoids an external FFmpeg process, according to the README, but the project warns that AVC and HEVC parsers can create patent obligations. Masking downloads a SAM checkpoint on first use; SAM 2.1 is Apache-2.0, while SAM 3 has Meta's separate license.
What happened when we ran it
Our sandbox cloned commit 3a972c9 and installed 35 Python packages in 18 seconds. They occupied 37 MB on disk. The detected build step returned successfully in 1 second. Pip-audit found 0 known vulnerabilities in those installed Python packages. The checkout itself contained 1,360 files and about 326,068 lines of source.
There was no test script or test target, so the lab skipped tests. That leaves an important limit on the result: we confirmed that the harness's install and build commands completed, but we did not get an automated correctness signal from the repository. The scan found 2 CI workflow files, no Dockerfile, and no tests directory. It also did not run a GPU reconstruction, so it says nothing about image quality, training speed, VRAM use, or mesh quality.
A low reprojection error can still hide a bad room
Issue 77 is the strongest reason to inspect the solved cameras before training. A user reported a 0.83-pixel average error while some cameras sat outside an apartment and a bed appeared above a toilet. The maintainer called this a known structure-from-motion failure mode, especially with low texture or repeated texture, and suggested trying different releases, capture formats, feature settings, frame rates, mappers, and masks.
That reply is useful because it does not pretend the score guarantees a correct scene. If you scan buildings or repetitive interiors, review the camera plot and a quick result before spending more time on a detailed run. The current release says it improves 360-camera solving with rig constraints, but its release note also asks users to report any accuracy or speed regression. Improvement here is active work, not a solved input problem.
Current reports make mesh export and Windows downloads judgment calls
Issue 91 records an access violation during baked-texture meshing on Windows 10. The report describes 207 EXR images, a 2.9-million-splat scene, and a crash after UV processing; it says the same operation only failed when baked texture was enabled. The issue was still open when we checked, so a team depending on mesh delivery should reproduce its own scene and settings before settling on this exporter.
Issue 92 needs equally careful wording. Two users said Windows Defender flagged the v2026.9.20 Vulkan ZIP as Trojan:Script/Wacatac.H!ml. The maintainer could not reproduce the detection and submitted the exact archive to Microsoft, while another commenter reported a clean VirusTotal result. That is not proof of malware or proof that every machine will accept the file. A managed Windows environment may need its security team to verify and approve the binary.
The project is moving quickly and depends heavily on one maintainer
GitHub showed 654 stars, 36 combined open issues and pull requests, and a last push on September 23, 2026. Release v2026.9.20 arrived 3 days earlier with adaptive frame extraction, warm-start training from a PLY file, presets, batch processing, GPU selection, RGBA training, and meshing fixes. Recent pull requests and same-day issue replies show real activity rather than a dormant release page.
The README also says Spirula Studio is developed and maintained almost entirely by one person. That concentration matters for a tool spanning camera solving, GPU code, model downloads, a desktop UI, and several operating systems. Spirula Studio is worth a trial because it removes a lot of assembly work. Its safest use starts with a small representative capture, a manual camera check, and a successful mesh export on the exact release you plan to keep.