Spirula Studio v2026.10.16 adds spherical beta, a view-dependent color model for gaussian splatting, as an alternative to spherical harmonics.
Spherical beta comes from Deformable Beta Splatting, published by Liu et al. at SIGGRAPH 2025. Each splat carries a few explicit specular lobes, each one six numbers for color, direction and sharpness, in place of or on top of its SH bands. The lobes only ever add light. The Gaussian kernel and the 3dgs, mip and 3dgut primitives are unchanged.
Spherical beta runs in Vulkan builds only, and the CUDA launchers refuse a model with lobes. The developer says exported files open in gsplat and PlayCanvas, though not in the Deformable Beta Splatting reference code's own reader.
Against SH 3, two lobes scored 31.57 versus 31.31 PSNR on bonsai, 29.55 versus 28.70 on counter, 27.07 versus 26.68 on garden and 31.36 versus 31.35 on room, with SSIM within 0.002. On bonsai, two lobes cut summed FPBO GPU time from 201 s to 59 s, all kernels from 539 s to 388 s, and splat VRAM from 351 MiB to 222 MiB.
Spirula Studio v2026.10.16 also adds dense initialization. A `spirula dense` command builds a colored point cloud from registered images with the camera poses held fixed, using RoMa v2, and training starts from that cloud. RoMa runs through the existing Vulkan inference layer with no Python, PyTorch or ONNX Runtime dependency. The weights are not bundled, and the GUI downloads them after showing the model terms. The feature, inspired by the LichtFeld Densification Plugin, came in a pull request from contributor D1odeKing.
The update follows v2026.10.6, which brought E57, LAS and PLY laser scans into training. New in this release are Metashape .psx parsing, a per-input .cube LUT baked in at frame extraction, better support for DJI D-Log M and Insta360 I-Log footage, IMU and GPS parsing from Antigravity A1 footage, screenshots and saved camera views, and source preview with video trimming. Merged models show fewer seams after partitioned training, introduced in v2026.9.30. RAM and VRAM use drops for SfM on large datasets and for high resolution training, though the release notes say SfM can be similar or slower depending on disk bandwidth, and that quality should remain near-identical. Many of the new features come from community contributors.
Spirula Studio v2026.10.16 is available now on GitHub under the GPL-3.0 license.
