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 build an efficient workflow.
In the previous article (Basic Workflow), I introduced the minimum viable workflow I use for Gaussian Splatting as a hobby, built entirely around free-to-use tools. This time, I'll explain how to further improve quality with just a few small tweaks to that workflow. Additionally, for a workflow using the paid software Agisoft Metashape, please refer to the "Metashape Edition" if you are interested.
💡 Before You Start
This article includes a step that involves running a Python script. If you're new to Gaussian Splatting or unfamiliar with Python, I recommend working through the Basic Workflow first to get a feel for the overall process before tackling this one.
Output Comparison with the Basic Workflow
The image below compares the output from the basic workflow and this article's approach. You can see that tile seams and object outlines are reproduced more clearly.
📝 Note
This article is written so it can be followed independently, so some sections overlap with the previous article. Please keep this in mind.
Differences from the Basic Workflow
Here's a summary of the main changes for those who have already read the previous article.
Basic Workflow (previous):
Generate SfM and a point cloud from Equirectangular images → Feed the Cube Map split images directly into the Gaussian Splatting GUI tool
Quality Improvement Edition (this article):
After generating SfM and a point cloud from Equirectangular images, the following steps are added:
1. Excluding top and bottom images — When splitting into multiple viewpoints, images showing mostly sky or ground are excluded. This reduces low-texture regions that can cause noise.
2. Point cloud regeneration — The point cloud is recalculated from the extracted images. By generating a point cloud that matches the pinhole camera model used during Gaussian Splatting training (rather than the Equirectangular model used during SfM), noise factors during training are eliminated.
⚠️ Prerequisites
This article assumes that the 360° camera is held nearly vertical during shooting.
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) |
| RealityScan | Used for point cloud regeneration |
| 360-to-RealityScan | Converts SphereSfM results into a format readable by RealityScan (.xmp) |
※ (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.
Other
-
Python environment (e.g., Anaconda)
Required only if you run 360-to-RealityScan directly as a Python script.
Not required if you use the pre-built executable (.exe) for 360-to-RealityScan. Download SphereSfMToRS.exe from Releases.
Overall Workflow
Gaussian Splatting generally follows these 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 |
In this article, a point cloud regeneration step is added between Step 3 and Step 4.
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
💡 Tip
With stabilization ON, the camera automatically corrects for tilt. Strictly speaking, the slight distortion introduced by this correction can become a source of noise in later processing — but if you just want to get started without worrying about that, turning stabilization ON is totally fine.
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 GUT
⚠️ Difference from the Basic Workflow
The Basic Workflow usesSphereSFM, but this article usesSphereSFM GUT. Please make sure to select the correct one.
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:
💡 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
└── 📁 lichtfeld
├── 📁 images
├── 📁 masks
├── 📄 pointcloud.ply
└── 📄 transforms.json
In preparation for the next step, create a RealityScan folder at the same level as the lichtfeld folder.
📁 Folder containing the video
└── 📁 Folder with the same name as the video
├── 📁 lichtfeld
│ ├── 📁 images
│ ├── 📁 masks
│ ├── 📄 pointcloud.ply
│ └── 📄 transforms.json
└── 📁 RealityScan ← folder you just created
Step 3. Point Cloud Regeneration
This is the key step in this article. The SfM results are imported into RealityScan to regenerate a point cloud suited to the pinhole camera model.
3.1 Input/Output File Configuration
1. Download SphereSfMToRS.exe from Releases and launch it by double-clicking (no Anaconda or Python required)
On first launch, enter the path to ffmpeg.exe in the FFmpeg (.exe) path field at the bottom of the window and click 💾 to save
[!NOTE]
Running as a Python script
Open a terminal (Python environment) and run:cd C:\path\to\360-to-RealityScan python spheresfm_to_realityscan.pyFor setup instructions, refer to the README at https://github.com/TakashiYoshinaga/360-to-RealityScan
2. Specify each path as follows:
| Item | Target |
|---|---|
| transform.json |
transforms.json inside the lichtfeld folder |
| PLY file (optional) |
pointcloud.ply inside the lichtfeld folder |
| Input folder (equirectangular) |
images folder |
| Mask folder (equirectangular, optional) |
masks folder |
| Output folder | the RealityScan folder you created |
3.2 Split Settings
The default settings are generally fine.
| Parameter | Description |
|---|---|
| Additional Pitch Angles | The equator (vertical center of the image = 0°) is always included automatically. To also crop directions tilted up or down from the center of the equirectangular image, enter angles as comma-separated values (e.g. -45,45 adds 45° above and 45° below center). Leave blank to crop the equatorial direction only |
| Split Count (Equator) | Number of splits along the equator. Default is 6; setting to 4 extracts only the side faces of the Cube Map |
💡 Note on Additional Pitch Angles
Depending on the scene, adding upward and downward views can improve the quality of the Gaussian Splatting output. For example, if the floor or ceiling has distinctive textures or patterns, try adding-45,45to include tilted angles.
3.3 Running the Conversion
Click Start Conversion to start the conversion.
When complete, you'll see the following screen:
The SfM results are now ready to be imported into RealityScan.
📝 Note
If you specify bothtransforms.jsonand the PLY file, the output will include complete COLMAP-format data containing camera poses, a point cloud, and image information.
If you prefer to skip point cloud regeneration in RealityScan and use the images split by this tool directly for Gaussian Splatting, proceed to Step 4. Gaussian Splatting.
3.4 Point Cloud Generation in RealityScan
1. Launch RealityScan
2. Click the WORKFLOW tab
3. Click Folder and select the all folder inside the RealityScan folder
4. Open Inputs and press Ctrl + A to select all images
5. Configure the detail settings as follows:
| Category | Item | Value |
|---|---|---|
| Prior pose | Absolute pose | Locked |
| Prior calibration | Calibration group |
0 (same settings for all cameras) |
| Prior |
Fixed (fixed field of view) |
|
| Prior lens distortion | Lens group |
0 (same settings for all cameras) |
| Prior |
Fixed (fixed distortion) |
6. Click the ALIGNMENT tab
7. Click Align Images
When processing is complete, the regenerated point cloud will be displayed.
3.5 Export
[Export Camera Pose Information and Point Cloud in COLMAP Format]
1. Click the WORKFLOW tab
2. Click Export
3. Click COLMAP
4. Save to the RealityScan folder with any file name (e.g., colmap)
📝 Note on File Overwriting
Thecameras.txt,images.txt, andpoints3D.txtoutput byspheresfm_to_realityscan.pywill be overwritten. This is fine, but if you want to keep the COLMAP files from before regeneration, move them to a separate folder beforehand.
5. In the Export Dialog, set Directory structure to Flat
6. Set Export masks to No
7. Temporarily set Export images to Yes
8. The Export image setting will appear
9. Set Image format to jpg
10. Set Naming convention to original file name
11. Set Export images back to No
12. Click OK
💡 Hint
Theallfolder and its images are used exclusively for processing within RealityScan. Therefore, you can safely delete them once the export is complete and they are no longer needed. If you ever need them again, you can always regenerate them by re-runningspheresfm_to_realityscan.py.
Step 4. Gaussian Splatting
4.1 Loading Data
1. Launch LichtFeld Studio
2. Drag and drop the RealityScan folder (directly under the video name folder) 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.
4.2 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.
4.3 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.
4.4 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 confirmed that adding just a few steps to the basic workflow can produce a clear improvement in quality. The key points are:
1. Reducing noise by excluding low-texture regions (top and bottom images)
2. Improving training accuracy by regenerating a point cloud matched to the pinhole camera model
Underlying both of these is a fundamental principle: high-quality output requires high-quality input.
- In SfM, the quality of the input images and the accuracy of the mask information directly determine how well camera poses can be estimated.
- In Gaussian Splatting training, the accuracy of the SfM results — along with continued use of high-quality images and masks — determines the quality of the final output.
Being mindful of input quality at each step is, in my view, the most reliable path to better results.
By building on this article's workflow with further adjustments, you should be able to pursue even higher quality Gaussian Splatting results.
Sample Output
The video below shows the result of applying this workflow to a wider area than covered in the Basic Workflow edition. The Basic Workflow approach failed to converge during training and did not produce a clean output, but this quality improvement method succeeded.
