Spirula Studio v2026.9.30 can split a reconstruction too large for one training run into parts, train them separately, and merge the results.
The GPL-3.0 application already carried structure from motion, AI masking and gaussian splatting training in one binary on NVIDIA, AMD, Intel and Apple GPUs, and v2026.9.24 added editing and an MCP server six days earlier. Until this release, every dataset trained as a single run.
A Partition button on the dataset screen, and the spirula partition command, cut the scene by visibility, so each camera sees mostly its own part and the cut is meant to follow walls rather than run through rooms. By default each part holds at most 2,000 images, counting a ring of outside cameras that have at least 10 percent of their view inside the part, so neighboring parts learn a seam from the same views. The merge assigns every splat to the single part that owns its position, with no feathering.
On a 4,102 frame capture of an open hall, a three part merge scored 22.63 dB PSNR against 22.73 dB for one model trained on everything. On a 7,013 frame capture of a building on five floors, trained at half resolution with a wider 20 percent ring, a six part merge scored 17.11 dB, the same as one model trained on everything. The design notes state that color seams from exposure drift between parts are not yet addressed.
Masking now takes natural language prompts for SAM 2.1 models, and the default changes from SAM 3 to SAM 2.1 Base+ with Grounding DINO, which the notes put at 2 to 3x faster with slightly reduced accuracy. SAM 3 remains available. BiRefNet joins as a masking model, alongside a pen tool and a separate fixed area mask for each camera in dual fisheye setups.
Spirula Studio v2026.9.30 forecasts training VRAM, warns of out of memory risk, and estimates time to completion more accurately. Dataset creation is faster and makes better use of telemetry, and both dataset creation and meshing are less likely to hit a "VkResult -4" error. The developer flags a structure from motion change as intended improvement and asks for reports of any regression. Fixes cover PLY files saved from resumed training, COLMAP datasets with relative paths, and video decoding on some drivers.
The commit history also adds complete calibrated camera intrinsics for structure from motion, co-authored by Jeff Powers of Arcturus Vision, which released Gaussian XR for Steam Frame on September 28.
