Video to Splat
Video to Splat allows you to create a Splat from a video or a set of images. This feature automatically converts visual input into a 3D point-based representation, making it easy to bring real-world scenes into Arrival.Space. You can use it to capture objects, environments, or short clips and turn them into 3D content without the need for complex 3D modeling.
Interface
To start the Video/Image to Splat process, use the Action Button in the bottom-right corner 1 and select Video to Splat 2.
3 Upload Files
The Video to Splat Interface will open, allowing you to choose whether you want to upload images or a video. To select your files, simply drag and drop them into the upload area, or click inside the field to open your system’s file manager.
Only Mobile
You can record a video directly with your phone and upload it without saving it locally. After the training has successfully completed, you’ll have the option to download the files you uploaded.
4 Presets
Presets are templates that automatically apply predefined configurations in addition to the Advanced Settings. If you don’t modify the advanced settings manually, only the preset values will be used. The following options are available:
| Name | Max Steps | Image Resolution | Number of Frames | SH Degree |
|---|---|---|---|---|
| Draft | 4k | 2k | 300 | Level 1 |
| Normal | 20k | 2k | 300 | Level 2 |
| High | 25k | 4k | 600 | Level 3 |
5 Create New Space
If this setting is disabled, the generated Splat will be placed inside your current Space. If you are not inside your own Space, you cannot change this setting.
6 Advanced Settings
The Advanced Settings allow you to modify the template configurations. These settings are currently work in progress, so their functionality is still limited.
After adjusting the settings, simply click the Back Button in the top-left corner next to the title, then press Submit to start the job.
1 Max Steps (k) (Splat Training)
Defines the number of training steps applied to the Splat. A higher value increases detail but also processing time.
2 Image Resolution (Camera Reconstruction)
Specifies the resolution of the undistorted images used for camera reconstruction.
3 Number of Frames (Frame Selection)
Sets the maximum number of frames to be used for generating the Splat.
4 SH Degree (Splat Training)
Controls the degree of spherical harmonics, which affects lighting and color accuracy.
5 Multi-Camera Support (Camera Reconstruction)
Enable this option when uploading multiple videos or when footage was captured using different cameras.
Process
As soon as you hit Submit the selected files will be uploaded and you can't leave until the files are uploaded. If you have a bad internet connection or you have to leave - don't worry. You have ~24hours time to upload the same files and continue uploading, but this only works if they have the same name.
The following Images represents the process while the Splat get's generated. It doesn't matter if it happens in a new space or as part of an existing space. The only difference is, that you get a new space with the file title - in case of multiple files the first file name.
Currently, you can only submit one Job at a time. If you try to submit a new Job, you will be asked if you want to abort the old one.
1 2 Progress bar
In the space where the splat will be placed, you can find a progress bar at the bottom. If you click/touch on it, the space Edit UI will open with the detail view of the Video to Splat progress. The progress itself has following Steps:
1. Initializing the Job
The system checks if the service is currently in use. If the server is already processing another splat, your job will be queued.
2. Extracting Frames (only for videos)
A large set of images is extracted from the uploaded video.
3. Calculating the Sharpness (only for videos)
The system evaluates all extracted frames and selects the sharpest ones for further processing.
4. Feature Extraction
Key visual features such as points, edges, and textures are detected and stored for each selected image.
5. Feature Matching
Corresponding features across multiple frames are matched to establish spatial relationships between images.
6. Image Undistortion
Lens distortion is corrected using the camera parameters to ensure accurate geometry for 3D reconstruction.
7. Splat Training
The 3D Gaussian splatting model is trained using the processed image data to reconstruct the scene in high quality.
3 Abort the Job
If you want to cancel the job, you can do so by clicking the icon.
4 Splat Preview
As soon as the system starts the Splat training, a preview will appear inside the Detail View.
5 Source Material
You can download your source files by using the icon. This option is especially useful when recording directly with a smartphone, as it allows you to review and evaluate your captured material.
6 Parameters
The Parameters section shows all applied settings. This is helpful if you want to reuse the same source material but adjust specific settings to improve your results.
Best Practice
To create a high-quality 3D model, there are a few important rules to keep in mind when capturing your material.
Recommended Capture Guidelines
- Move the camera in a circular path around the object to cover it evenly from all sides.
- Keep the object in the center of your camera view.
- Change your point of view but don't pan, tilt or roll your camera.
- Capture from as many different angles as possible — including higher and lower viewpoints.
- Lock your focal length to maintain consistent field of view across all frames.
- Use a fixed ISO setting to ensure uniform exposure and prevent brightness shifts.
- Set manual focus so you can precisely define the area or object of interest.
- If possible, use even, diffuse lighting (e.g., cloudy daylight or soft indoor lighting) for optimal texture capture.
Images
When capturing images instead of video, take at least 50 photos, ideally up to 200, depending on the object’s size and complexity.
Photograph the object from different sides and angles, including top and lower views.
Use manual focus and a fixed ISO.
Videos
When recording a video for 3D reconstruction, keep the duration between 20 and 120 seconds.
Move the camera smoothly in a circular path around the object and capture it from multiple angles and heights.
Avoid quick movements, strong light contrasts.