Every 360 Gaussian Pro release, newest first. The app checks for updates on start — see Installation if you'd rather grab the installer yourself.

1.2.0 — 19 October 2026

AMD and Intel GPU support: the whole pipeline (extract, mask, align and train) now runs without an NVIDIA card.
Masking works on AMD and Intel GPUs (ONNX + DirectML).
SIFT features now run on any GPU through Vulkan. SiftGPU has been removed.
Brush is bundled as the default trainer, and its window opens inside the Train tab.
Rolling shutter correction, on by default, for Insta360 X4 / X5 / X6, Antigravity A1, DJI Osmo 360 and Avata 360.
GoPro videos (experimental): gravity and GPS are read from the clip, so the model comes out upright and to scale.
HDR training data: Insta360 X6 I-Log .insv clips export linear EXR tiles for HDR training in Postshot.
Camera filter slider in the 3D preview: align with all frames, then train with only a share of them for faster training.
Optional .txt model export.
Command-line version (experimental): 360GaussianPro-cli.exe runs extract, mask, align, mesh and train without opening the window.
Log view: Ask AI explains errors and can send the log to the developer.
Project folder: working files now go into a pipeline/ subfolder. The project root keeps only the training data, mesh, training output and log.
COLMAP updated to 4.2.1.
Updated for the LichtFeld Studio 0.5.4 command line.
Fixed: masking ran out of memory on long clips and large batches (thousands of frames). Masks are now saved as they're made, so a stopped run resumes where it left off, and an out-of-memory error stops the step instead of looping.

1.1.1 — 5 October 2026

Sharper distant surroundings: far-away hills, trees and buildings are no longer smeared in the trained splat.
Each step now shows an estimate of how long it will take.
Fixed: tilted ground on .insv videos longer than ~20 seconds (motion sensor data was cut short).

Masking tab: new picker mode.
Simple mesh generation.
Version rollback.
Denser alignment: two-view points are now kept, giving 2.7× more points and lower reprojection error with the same runtime.
Lens calibrations are only reused when the new solve measures better than the stored one.
GPS scaling now checks the shape of the GPS track against the camera path and ignores bad fixes.
Split images carry their lens data (focal length) in EXIF, ready for RealityScan and Metashape.
If the video has GPS, every extracted frame and split image gets its position written to EXIF.
Smoother extraction timeline preview.
.insv videos and .insp photos can be batched into one project.

Initial release.

Fixed install and downloads on networks or antivirus software that block r2.dev.
A custom output path now moves all the files, not just the training output.
Postshot Photometric Compensation option added to the UI.