Introduction
Gaussian Splatting has been gaining attention as a method for recreating wide outdoor scenes in 3D. While related information is increasingly appearing on social media and blogs, it is scattered across many sources, making it difficult to piece together an efficient workflow. The process also requires switching between multiple applications, which can feel daunting for beginners.
In this article, I'll introduce the minimum viable workflow I personally use for Gaussian Splatting as a hobby.
All tools covered here are free to use (with some optional paid features), making it easy to get started even if you're new to Gaussian Splatting.
For techniques to further improve quality, please refer to the companion article "Quality Improvement Edition". Additionally, I have introduced another workflow using the paid software Agisoft Metashape in the "Metashape Edition". If you are interested, please check that out as well!
Sample Output from This Workflow
Environment
PC Specs
| Item | Spec |
|---|---|
| OS | Windows 11 |
| GPU | NVIDIA GeForce RTX 4070 SUPER |
| CPU | AMD Ryzen7 8700G |
| RAM | 32 GB |
360° Camera
-
Insta360 X4 Air
※ Other manufacturers' cameras such as THETA are also supported. 8K or higher is recommended.
Software
| Software | Purpose |
|---|---|
| LichtFeld Studio v0.5.2 | GUI tool for 3D Gaussian Splatting ※ |
| 360° Gaussian v1.4.3 | Tool to automate each step of Gaussian Splatting (Updated: April 8, 2026) |
※ (Updated: April 18, 2026) Free distribution of pre-built binaries for LichtFeld Studio v0.5.2 has ended. v0.4.0 remains available for free download. To use v0.5.2, you must either purchase a paid license or build it from source.
If you want to try the latest features available on GitHub, or try v0.5.2 before purchasing a license, please refer to the notes below and try building it yourself.Notes for building from source
Click to expand
To build it yourself, follow the build instructions here.
Tip 1
Also note: although not mentioned in the official build instructions, Perl may be required. Install it from Strawberry Perl.
Be aware that the cmake bundled with Perl may take precedence over your system cmake, so first temporarily renameC:\Strawberry\c\bin\cmake.exe, and delete it only if necessary. To be safe, restart your PC afterward.
Tip 2
In the official documentation's Clone repository section, the following command did not always download all required files correctly in the author's environment:
git clone https://github.com/MrNeRF/LichtFeld-Studio
Instead, to avoid missing submodule downloads, it is recommended to add the--recursiveflag, or use a client tool such as Fork or GitHub Desktop:
git clone --recursive https://github.com/MrNeRF/LichtFeld-Studio
Tip 3
Note: skip thegit checkout v0.4.0command in the Checkout stable version step.
The build process takes a long time, so plan to run it when you have plenty of time available.
📢 Updating 360° Gaussian
When you launch 360° Gaussian, if “update” is displayed in the lower-right corner of the screen as shown below, make sure to update the application before proceeding.
If you have the optional AutoMasker app installed, also check whether an update is available for AutoMasker as well.
Overall Workflow
Gaussian Splatting generally follows these 4 steps:
| # | Process | Description |
|---|---|---|
| 1 | Capture | Shoot the scene with a 360° camera |
| 2 | SfM (Structure from Motion) | Estimate the position from which each image was taken |
| 3 | Point Cloud Generation | Generate a point cloud based on the camera positions from SfM |
| 4 | Gaussian Splatting | Generate a 3D Gaussian Splatting model from the point cloud |
Step 1. Capturing & Exporting the Video
Capture the scene with your 360° camera. Since processing can take a long time, it's recommended to start with videos under 1 minute until you get familiar with the workflow.
Export the captured video in Equirectangular format and copy it to your PC.
If you're using an Insta360, export it using Insta360 Studio (the PC application).
⚠️ Note on Stabilization
Stabilization can be ON or OFF, but if you turn it ON, make sure to disable the following:
- Horizon Leveling
- Tilt Recovery
- Vibration Reduction
Step 2. SfM and Point Cloud Generation
This step uses 360° Gaussian.
Because 360° Gaussian supports multiple tools for both SfM and Gaussian Splatting, it makes it easy to compare and test different combinations of tools and methods.
For detailed usage, refer to the following videos:
2.1 Image Extraction
1. Launch 360° Gaussian
2. Click Add Video(s) and select the Equirectangular video
3. Select Splitting and configure the extraction settings
| Parameter | Description |
|---|---|
| Extra frame every | Extract images at the specified interval (seconds or frames) |
| Sharp frame extraction | Whether to prioritize less blurry images compared to adjacent frames |
| Sharpness check range | E.g., 10 compares ±5 frames to select the sharpest image |
2.2 Image Masking (Optional)
AutoMasker is a tool that automatically masks regions that are unnecessary for Gaussian Splatting.
ℹ️ This step is optional
If you are not using AutoMasker, skip ahead to 2.3 SfM Configuration.
1. Click AutoMasker
2. Enable Use AutoMasker
3. Enter keywords in Detection Keywords separated by periods (.)
Example: person.sky
💡 Two Reasons Why Masking Matters
First, subjects that move or lack distinct features — such as people, vehicles, and low-texture objects — introduce noise into the SfM process, reducing the accuracy of camera pose estimation. Since SfM accuracy directly impacts the quality of the Gaussian Splatting output, masking these subjects out is essential.
Second, the same noise degrades quality during the training stage as well. When the same subject appears with a different appearance across frames, the Gaussians struggle to converge to the correct shape and color. Masking protects quality at both stages of the pipeline.
📺 Reference Video
📝 About AutoMasker
AutoMasker is a paid tool (46€), but since it runs as a standalone application, it can be used not only within the 360° Gaussian Splatting automation workflow, but also for general 360° image masking tasks outside of Gaussian Splatting.
It offers better value than purchasing a similar tool, and is well worth considering.
For setup instructions to integrate it with 360° Gaussian after purchase, refer to this video.
2.3 SfM Configuration
1. Click Alignment
2. Configure Training Method
Selecting LichtFeld will run SfM, point cloud generation, and Gaussian Splatting all at once, but the LichtFeld Studio window won't open during processing, so you cannot observe the progress.
My preferred settings are:
-
Training Method:
No Training -
SfM:
SphereSFM
⚠️ Do not select
SphereSFM GUT
The workflow usingSphereSFM GUTis covered in the companion article "Quality Improvement Edition". In this article, selectSphereSFM(without the GUT suffix).
3. Open Advanced SphereSFM Setting and select the camera image size from Preset
Example: If shooting in 8K, select Ultra (8K)
4. Select sequential for Matcher
Since v1.4.0, Loop detection with vocabulary tree is now supported, significantly improving loop closing performance.
Other settings can be left at their defaults.
As a reference, below is a practical settings example that the author often uses. In scenes that include low-texture areas, enabling Realign cubemaps can sometimes introduce misalignment, so the author often keeps it disabled. Please choose the setting based on the scene:
💡 If sequential doesn't work: Try exhaustive as a last resort (Click to expand)
If sequential doesn't produce good results, exhaustive is an option to try as a last resort. It matches all image pairs exhaustively, which significantly increases processing time, but may cover pairs that sequential failed to connect.
As an example, here are the settings the author typically uses for shooting relatively large areas:
Note: This configuration increases the number of computation iterations compared to the default settings and may take more time. We recommend starting with the default settings first.
The author's recommended configuration strategy is as follows:
-
Iterations(Global) and Refinements (Always High)
- Setting these high makes it easier to find the correct geometry even for difficult image pairs with a low inlier ratio.
-
MinInliers (Minimum Inlier Count)
- Large-scale areas (Low): Keep this low. This allows pairs with few feature points—often judged as "weak matches"—to pass verification more easily, ensuring you don't miss the slight clues needed for loop connections.
- (Note) Small areas (High): Since feature points match consistently across images, setting this intentionally higher ensures only "reliable matches" are adopted. This eliminates erroneous matches and results in higher accuracy.
-
Reprojection threshold
- Large-scale areas (Loose): Changes in lighting, weather, and perspective can cause major viewpoint differences between the start and end of a loop. If the threshold is strict here, sparse matches will be rejected, preventing the loop from closing (connecting). Therefore, easing the tolerance to match the scene's variability is the standard approach.
- (Note) Small areas (Strict): Because viewpoint changes are minor and feature points are stable, you can afford to use a stricter (tighter) threshold. This helps eliminate noise and allows for higher precision camera pose estimation.
In short, for large-scale scans, the combination of "verify thoroughly (High Iterations(Global)/Refinements), don't discard the results (Low MinInliers), and ease the tolerance in response to spatial diversity (Loose Reprojection threshold)" is designed to maximize the success rate of SfM.
2.4 Running SfM
Click Start to run SfM.
When processing is complete, you'll see the following screen:
💡 Note for first-time runs
On the first run, processing may complete instantly. In this case, Failed to process: may appear before Batch Processing Complete.
If this happens, click the Stop button once and try running it again.
Verify that the camera positions and point cloud have been generated correctly.
📁 Folder containing the video
└── 📁 Folder with the same name as the video
└── 📁 train_data
├── 📁 images
├── 📁 masks
└── 📁 sparse
Step 3. Gaussian Splatting
3.1 Loading Data
1. Launch LichtFeld Studio
2. Drag and drop the train_data folder generated above into the window
3. When the Load DataSet dialog appears, click Load
Verify that the point cloud and images have been loaded correctly.
If you don't need to see the camera images, uncheck Camera Frustum in the Rendering tab on the right side of the screen
⚠️ Note
If the images are still loading, some faces of the Cubes might not display the images. Starting the Gaussian Splatting training in this state may cause LichtFeld Studio to crash.
3.2 Training Configuration
Here is an example training configuration. Feel free to experiment with different settings as you get more familiar.
1. Click the Training tab
2. Select MRNF for Strategy
3. Set Steps Scaler appropriately
| Condition | Recommended Value |
|---|---|
| 300 or fewer images | 1 |
| More than 300 images | number of images / 300 |
⚠️ If Training Doesn't Work
If the Gaussian Splatting training fails to converge and the view whites out, setting Steps Scaler to 2–3x the value ofnumber of images / 300tends to stabilize training.
4. Set Max Gaussians for the maximum number of Gaussians
The default value is generally fine, but increase it if the output lacks detail.
Optional Settings:
Only apply the following if you created mask images with AutoMasker.
-
Mask Mode → Set to
Ignore - Use Alpha as Mask → Uncheck
For other parameters, start with the above settings and experiment as you become more comfortable.
3.3 Running Training
1. Use the mouse to zoom in on the area you want to observe during training
(In my example, around a bridge)
2. Click Start Training to begin
3. The display starts blurry but gradually becomes clearer as steps progress
Training stops automatically when the step limit is reached.
If you want to save intermediate results, click Save Checkpoint.
3.4 Export
The exported data can be used with tools like SuperSplat Editor for creating videos or viewing in a viewer.
1. Click File → Export
2. Select the output format (e.g., .ply)
Summary
Through testing with various indoor and outdoor scenes, I've found a workflow that produces reasonably stable quality results. By tweaking the settings in this article, you should be able to achieve even higher quality Gaussian Splatting.
For an approach that takes things a step further, details are covered in the companion article "Quality Improvement Edition", so please check it out if you're interested.