v2026.10.16

Changelog
- Optimized RAM and VRAM usage for SfM on large datasets
- Optimized VRAM usage when training on high resolution images
- Support dense initialization, inspired by [LichtFeld Densification Plugin](https://github.com/shadygm/lichtfeld-densification-plugin)
- Support taking screenshots and saving camera views during training and viewing
- Better support for creating datasets on log footage (DJI D-Log M, Insta360 I-Log, etc.)
- Improved use of GPS in dataset creation
- Support source preview and video trimming in dataset creation
- Support spherical beta from [Deformable Beta Splatting](https://rongliu-leo.github.io/beta-splatting/) for view-dependent appearance
- Reduced chance of watchdog timeout error (VkResult -4 / "Impacting Interactivity") on slower GPUs
- Reduced seams in merged models after partitioned training
- Revised masking prompts for dataset creation with 360 cameras
- Made scene centering automatic for datasets in ECEF-like frames to reduce precision issues
- Support parsing IMU and GPS from Antigravity A1 footage

Notes
- This release carries RAM and VRAM optimizations in SfM and training. Expect similar training speed, SfM speed can be similar or slower depending on disk bandwidth. Quality should remain near-identical. Please submit an issue if you notice significant regression compared to previous versions on your dataset.
- A lot of great new features in this release are made by community contributors. Thanks those who contributed or tried to contribute, and we continue to welcome contributions, no matter large or small.