Align and Export 3DGS Cameras for D5 Render with RealityScan

What will you learn in this tutorial?

This guide covers aligning cameras from video frames or photos in RealityScan and preparing the results for D5 Render’s 3D Gaussian Splatting (3DGS) workflow. By the end, you’ll have a folder containing the source images, a JSON file with camera poses, and a PLY file containing the initial sparse point cloud. These three assets will be imported into D5 Render’s 3DGS creation window.

The goal isn’t to achieve the same number of cameras or the same quality of results in every RealityScan project. Image coverage, capture quality, and the components generated during alignment can vary from project to project. That’s why it’s important to select the right component and check the results before exporting.

Requirements: software, account, and source files

The workflow uses the desktop version of RealityScan and D5 Render’s 3DGS creation window. The D5 Render guide explains how to install and launch RealityScan through the Epic Games Launcher. You’ll need to sign in to the Launcher with an Epic Games account; review the pricing plan displayed in the interface and the End User License Agreement during installation as well. Since the source material doesn’t specify a price, refer to the terms shown on the installation screen.

In the Launcher, select the version to install from the RealityScan tab. The guide recommends the latest stable version selected by default for first-time users, but doesn’t require a specific release. Epic Games’ RealityScan 2.1 documentation also describes new features such as importing point clouds from SLAM scanners and using LiDAR classifications. However, the D5 Render workflow below follows the desktop features described in the source guide.

You can use either a video or a prepared photo folder as input. If you’re extracting video frames, decide where to save the video and frame folder in advance. Changing the frame file paths later may disrupt the subsequent loading steps described in the guide. The sources don’t specify any special hardware requirements for this process.

Step by step: align and export files

1. Start a clean new project. Launch RealityScan, go to the WORKFLOW page, and check that there are no inputs or components left over from an older project. Saving the project before you begin is a safety step recommended by the guide.

2. Import the video or images. You can drag a video file into the interface or use the Images / Folder buttons. Selecting a video opens the frame extraction window. The example in the D5 Render guide creates a frame folder alongside the video and sets Jumps’ length(s) to 0.5 seconds. In this example, that interval extracts one frame every 0.5 seconds; the appropriate interval may vary for different captures. If you’re using photos directly, select the existing photo folder.

3. Keep the frames and source folder in place. Don’t close the application while the video is being processed. Once processing is complete, check that the frame sequence appears in the Inputs panel. Don’t move the source video or frame folder during processing; you’ll need to point D5 Render to this folder in a later step.

4. Review the images. Open a few frames in the 2D view and check their sharpness, exposure, and coverage of the scene within the frame. The guide recommends avoiding motion blur, reflections, large textureless surfaces, and sudden exposure changes. Avoiding frequent zooming during video capture is another preparation tip from the source.

5. Align with Align Images. Run the Align Images command from the top toolbar. This detects features in the images, matches the frames, and calculates camera poses; it also generates sparse scene points. Allow for longer processing times as the number and resolution of the images increase.

6. Check the components. Once processing is complete, inspect the camera path and sparse points in the 3D view. The guide’s criteria include seeing Registered status on most images, a continuous camera distribution, and sparse points that describe the target scene. If multiple components are generated, don’t automatically assume that the one with the most registered cameras is the right one; also check its geometric consistency.

7. Select the valid component for export. The camera JSON and PLY file must come from the same component. Files exported from different components may have mismatched coordinate systems, camera counts, and point cloud positions. In the guide’s example, Component 1 is the largest, with 107/121 cameras; these figures apply only to the example project and aren’t expected values for every scene.

8. Save the camera transformation JSON. In the Export window’s Camera Alignment section, check Radiance Fields Transformation File and name the file transforms.json.

9. Export the sparse point cloud as PLY. Reopen the Export window and export Sparse point cloud under Sparse Scene Structure in Polygon File Format (*.ply). The guide uses sparse_points.ply as an example filename; use a different name if the next tool requires one.

10. Load the three inputs into D5 Render. Open the Create 3DGS from images/video window, enter a project name, and select Photos under Source Type. In the Photos section, choose the source photo folder or the folder containing the extracted frames. Under Advanced Training Files, add transforms.json for Camera poses and sparse_points.ply for Initial sparse point cloud. After confirming that all three files come from the same alignment component, submit the job using the Start reconstruction button.

Common mistakes and their impact on the workflow

The most critical mistake is exporting the JSON and PLY from different components. Verify the selected component before exporting and create both files from that same selection. Another potential issue is moving or renaming the video frame folder; the folder you load into D5 Render needs to remain accessible throughout the project.

If alignment is poor, first check the sharpness and coverage of the source images, as well as the overlap between adjacent frames. Avoid large jumps between viewpoints. If multiple components are generated, don’t decide based only on camera count; the cameras’ 3D distribution and the sparse points must also be consistent with the scene.

For architecture, interior design, and archviz teams, this workflow offers a documented way to take an existing space from video or photos into a 3DGS workflow. It may be useful to teams that want to document a space with images and send that data to D5 Render for reconstruction. However, beyond processing time, the sources provide no comparative information on hardware performance, licensing costs, or compatibility between specific versions. Verify these points separately against the installation screen and the versions you plan to use before starting a real project.

Next steps

Before submitting, double-check that the folder, JSON, and PLY are stored together and that all the files belong to the same alignment component. For those looking to automate the workflow, the RealityScan command-line documentation explains how to import video frames into a project with the importVideo command and run command sequences. If using command-line steps, verify the parameters in the relevant command documentation; don’t assume they replace the GUI-based export and loading process described in this guide.

Sources and licensing

This tutorial has been rewritten and adapted based on Epic Games’ RealityScan 2.1, project and image commands, and all commands documentation, as well as D5 Render’s guide to camera alignment with RealityScan. The source information is marked as free to use, but the supplied materials don’t specify a particular open license; no license name has therefore been stated.