Python plugins for LichtFeld Studio

Camera-path video recording for Gaussian Splat scenes.

Enables recording camera-path videos around Gaussian Splat scenes with linear and orbit segments inside LichtFeld Studio.

Convert equirectangular 360° video into masked COLMAP datasets for 3DGS training.

Automated pipeline for extracting and masking frames from equirectangular video, reframing them into pinhole perspectives using presets, and importing into LichtFeld Studio for training.

Select and enable camera subsets for training.

Provides camera selection tools for LichtFeld Studio, including random percentage selection, every-N selection, distinct camera filtering, crop-box visibility filtering, and controls for enabling, disabling, inverting, or hiding disabled cameras.

Run sparse COLMAP SfM reconstruction directly inside LichtFeld Studio.

Provides an integrated interface for running COLMAP-based Structure-from-Motion pipelines, including feature extraction, matching, and sparse reconstruction with quality metrics and live progress feedback.

Dense initialization plugin for LichtFeld Studio.

Performs a densification preprocessing pass on sparse reconstructions to generate denser point clouds inside LichtFeld Studio.

Depth-map generation and geometry-oriented inspection inside the studio.

Adds depth-oriented analysis and geometry inspection workflows directly to LichtFeld Studio for evaluation, debugging, and downstream tooling.

First-person ← → arrow key customisable navigation for LFS viewport. Includes 7 saved view slots with read/write of settings.

View navigation as First-Person with customisable floor (eye wont go below this set level). Plugin has eye/target XYZ display & allows 7 customisable view slots for your favoutite view angles: works in both viewer & Training modes. Settings persist on close so can be used in multiple projects.

One-click linked PLY sync from LichtFeld Studio to a Blender-oriented workflow.

Adds a one-click sync button for overwriting linked .ply files directly from LichtFeld Studio. Designed for workflows where GauSpla in Blender auto-reloads the same linked files for near real-time iteration.

Registers a scene to real-world GPS coordinates using EXIF, a similarity transform file, or an image positions CSV.

Solves a similarity transform (scale · R · t) that maps scene-space positions to ECEF (Earth-Centered Earth-Fixed / WGS-84) coordinates. Supports three input modes: automatic GPS extraction from image EXIF tags, loading a pre-computed similarity JSON, or loading an image positions CSV. Once registered, clicking any point on the model returns its latitude, longitude, and altitude.

Feed-forward 3D reconstruction from images, video, or COLMAP via Tencent HunyuanWorld-Mirror 2.0.

Runs Tencent HY-World-Mirror 2.0 on an image folder, video file, or existing COLMAP workspace and pushes Gaussian splats + camera poses + a point cloud straight into the LichtFeld scene — ready to view, train, or export. Supports direct in-memory trainer hand-off (no disk I/O), bf16 inference, calibrated VRAM auto-fit, and an image-subset picker for stride-sampling large frame folders.

3D measurement tools for Gaussian Splat scenes.

Measures distances between picked points on Gaussian Splats and shows the result directly inside LichtFeld Studio.

SHARP 4D media-to-Gaussian workflows for LichtFeld Studio.

Converts videos into animated Gaussian sequences or single images into Gaussian scenes inside LichtFeld Studio using SHARP.

Quick one-click overwrite of linked .ply files directly from LichtFeld Studio.

Adds a lightweight one-click sync button for overwriting linked .ply files directly from LichtFeld Studio without requiring a Blender-oriented workflow.

Separate outliers splats from the denser cloud. Inclues COLMAP editing/saving

Statistical Outlier Removal for 3DGS point clouds (ported from PointNuker). Two ways of editing COLMAP: (1) quick Crop & filter COLMAP sparse points in bin/txt or (2) GUI inferface with align, scale, rotation & cropping tools

Generate a 3D Gaussian splat from one or more images using QuerySplat and VGGT-Omega.

Runs feed-forward multi-view QuerySplat reconstruction from selected images, image folders, or clipboard images and inserts the resulting Gaussian splat directly into the LichtFeld scene. Local CUDA inference with background model preloading and VRAM release controls; approximately 9.1 GB of checkpoints download automatically on first use. VGGT-Omega requires approved Hugging Face access, acceptance of its FAIR terms, and local hf auth login; its weights are restricted to non-commercial research use.

Text-prompted object segmentation for sequential COLMAP datasets, generating per-frame binary masks for mask-aware Gaussian Splatting training.

Type a text prompt ("person", "red car", "action figure") and the plugin writes one binary mask PNG per frame into your dataset's masks/ folder. LichtFeld picks them up automatically — set Mask Mode to segment and only the prompted object survives training. Uses Grounding DINO (text → bounding box on frame 0) and SAM 2 (box → mask propagated across all frames). Models are cached in GPU memory after first load so repeat runs skip the 15–30s load step.

VR-oriented viewing workflow for immersive scene inspection.

Extends LichtFeld Studio with VR-focused viewing so scenes can be inspected and shared in a more immersive way.

Monitor Lichtfeld Studio training privately from Telegram.

Provides a private, read-only Telegram bot for checking training status, requesting camera renders and ground-truth comparisons, viewing charts and system health, and receiving error alerts without exposing an inbound port or enabling training controls.

Convert video to COLMAP datasets ready for Gaussian splat training.

Automated pipeline: extract frames from drone/camera video with GPS embedding from DJI SRT files, run 3D reconstruction via COLMAP, Agisoft Metashape, or RealityScan, and import the dataset into LichtFeld Studio for training.

Upload Gaussian splat scenes directly from LichtFeld Studio to SuperSplat.

Exports all or visible splats as PLY or SOG and uploads them to SuperSplat with progress reporting, resumable multipart transfers, and direct links to continue in the editor or viewer.

Generate a 3D Gaussian splat from a single image using VAST-AI TripoSplat, inserted directly into the scene.

Takes a single image, removes the background, and runs VAST-AI TripoSplat (single-image flow-matching) to produce a 3D Gaussian splat that is inserted directly into the LichtFeld scene. Guided three-step workflow (choose image, generate, place) with interactive gizmo placement that follows scene selection, adjustable detail / seed / guidance, and append-to-scene support. Local GPU inference; model weights (~3.8 GB) download on first use from Hugging Face.

Generate a 3D Gaussian splat from one perspective, panorama, or fisheye image using Insta360 UniSHARP.

Runs the UniSHARP single-image inference path inside LichtFeld Studio for perspective, panorama, and fisheye inputs, with optional camera calibration, plugin-local checkpoint/source downloads, and direct scene insertion through LFS scene.add_splat when available. Local CUDA inference; the UniSHARP checkpoint, UniK3D, and 3DGEER source dependencies are prepared on first use. Fisheye mode requires building the 3DGEER CUDA rasterizer with a CUDA-compatible compiler toolchain.

Feed-forward camera and point-cloud reconstruction from images or video using VGGT-Omega.

Runs VGGT-Omega on an image folder or video file and pushes predicted camera poses plus a point cloud directly into the LichtFeld scene, with optional COLMAP workspace export and train-from-points initialization. Non-commercial / research use only; requires approved Hugging Face access to facebook/VGGT-Omega and local hf auth login.

Export the current viewport as a JPG image at 1080p, 4K, or 8K resolution.

Capture exactly what the viewer shows and save it as a high-quality JPG. Supports 1080p, 4K, and 8K output heights with automatic aspect-ratio preservation and adjustable JPG quality.