Guide
Import your inputs, configure a reconstruction, train and manage local 01Splat projects.
Quick start
- Import your source
Start with an ordinary video or a photo set. Use 360° recordings for environment captures, a single photo for 2.5D, or import prepared COLMAP data with its original images. Check the import summary before continuing.
- Choose your settings
Choose a reconstruction type and, for multi-view data, a training profile. Review image size, splat limit, stopping steps and whether to keep training context.
- Generate or train
Start from the source to create a linked reconstruction object. Multi-view captures use iterative training; a single photo generates a 2.5D scene. You do not need to save a project first.
Input data
The input determines the preparation route. You do not need to build or connect processing nodes.
- Video and photos
- Import an ordinary video for frame extraction, or use a photo set. The route prepares the images and aligns cameras before Gaussian training.
- 360° environment captures
- Use 360° video or supported Insta360 recordings (.insv) for environment reconstruction. These inputs have separate preparation routes. Automatic 360° preparation requires Pro.
- Single photo · 2.5D
- Generate a 2.5D scene from one photo. This route does not use multi-view camera alignment or iterative training.
- Prepared COLMAP dataset
- Export camera calibration, registered image poses and sparse points in COLMAP format from your photogrammetry software. RealityScan and Metashape 2.2 or later export COLMAP directly; Meshroom needs a conversion through AliceVision's COLMAP export. Import the model and associate its original image files. A solved dataset skips camera alignment.
Before training, resolve missing or ambiguous image associations and any camera-model validation errors. Keep the original images and calibration intact. Input availability depends on the installed version.
Capture tips
- Use clear, overlapping views from different positions around the subject, rather than only turning the camera from one position.
- Keep exposure and white balance consistent where possible.
- For a static reconstruction, avoid motion in the scene. Use people masking to remove passers-by, not when a person is your subject.
Scene organization
Sources, reconstruction objects and folders share one scene tree. Organization and source relationships are separate.
- A source keeps the connection to its input media. Starting a new generation or training run from a source creates a new reconstruction object below it.
- To compare different settings, start another reconstruction from the same source. Each result has its own settings and training history.
- You can move a result to the scene root or a folder without changing its source relationship. Select a source or result to identify the related objects.
Reconstruction type
- General
- Multi-view reconstruction for objects, portraits or environments. It is not restricted to object captures.
- Environment
- Environment reconstruction with compatible geometry guidance. The selected training configuration must support the constraints you enable.
- 2.5D
- Single-photo reconstruction without multi-view camera alignment or iterative training.
Training profiles
Choose a profile in the selected source or reconstruction object's training settings. For your first multi-view run, start with MCMC.
- MCMC
- General-purpose training for objects, people and environments. Use it as a baseline when comparing results. Included in Free and Pro.
- Mirror
- Use for detail-oriented reconstruction, for example objects, portraits or textured surfaces. Requires Pro.
- Tron
- Choose this to use supported geometry constraints, especially for environments. It can also be used for objects with compatible geometry. Requires Pro.
Content type suggests a starting point; it does not lock the profile. Available choices depend on your license, GPU and input data. If a workflow requires geometry constraints, choose a profile that uses them.
To compare profiles, start another reconstruction from the same source. Resume continues the existing run and its original profile.
Training and preview
Settings belong to the selected source or reconstruction object, not to a global training tab. Review them before starting a run.
Main settings
- Limit image resolution
- Enable the limit and enter the longest image edge in pixels. For example, 1920 reduces a 3840 × 2160 image to 1920 × 1080. Lower resolution uses less GPU memory but may lose fine detail. Disable the switch to remove this limit.
- Maximum splat count
- An upper bound, not a target. When the field uses M splats, 2 means 2,000,000 splats. A higher cap can increase GPU memory use and project size.
- Stop training after
- Check the unit next to the field: in k steps, 30 means 30,000 steps. Start with the default and inspect the result before extending the run.
- Keep training context
- Enable this before starting to retain the complete state for later continuation. Saving a project with that state takes more space. This is Gaussian training state, not earlier capture preparation or camera reconstruction.
Inspect the reconstruction
Frame the selected object to find it in the viewport. Double-click a visible point on the reconstruction to set the orbit center.
The viewport remains navigable during training. Display controls show or hide splats, sparse points, capture cameras, axes and the grid. Hiding a layer or pausing live display updates does not pause the training task.
Pause, resume and reset
Request Pause and wait for the app to confirm that training has stopped and its complete state is saved. Resume continues this object; Reset starts over within it. To keep the old result for comparison, start another reconstruction from the source instead.
Scene undo and redo apply to scene edits, not to elapsed training or start, pause, resume and reset operations.
Projects and working data
Save a project whenever you need to. Importing, generating and training do not require a previously saved project.
The .01splat project records the scene, reconstruction results, settings and any retained training context. Original source media remain external; saving a project does not bundle those originals.
The global work folder holds computation data and can grow substantially. Choose a drive with enough free space and use the application’s storage controls to review working data rather than deleting active task folders manually.
Keep original media in their existing locations. If you move or rename them, their saved references may need to be relinked. A scene-tree move does not move the original files.
Export and continue later
Export a scene when you need a result for another viewer or application. Keep the .01splat project for scene organization and future edits.
A PLY result is not a complete training checkpoint. To continue the same training run later, retain its training context and required input files.
4D Gaussian Splatting is upcoming. A video used to reconstruct a static scene is still 3D; it does not by itself reconstruct motion over time.
Account and help
Sign-in is required to use 01Splat, including the Free plan. Account and subscription management are available from your account page.
To try Pro, start the 7-day Pro trial in 01Splat. It needs no payment method, is for evaluation only, and there is one trial per person, account, email address and device (a Windows user account on a computer). The Account menu shows when it ends. Afterwards, projects created during the trial open read-only: you can view them and export standard Gaussian splats; training, editing and saving need Pro.
For support, include the application version, input type, operation and error message. Do not send your source media unless it is necessary and you intend to share it.