How to Create a Perfect Gaussian Splatting Dataset with Blender and Synthra

If you've ever trained a Gaussian Splat from real photos, you know the frustrations: motion blur, inconsistent lighting, gaps in coverage. The training either fails or produces a blurry mess in the problem areas. Synthra, the Blender addon built by Luxa, solves this by generating the dataset entirely inside Blender: perfect images, precise camera positions, no surprises.

Why synthetic datasets work better for certain scenes

Real-world capture is great for spaces you can't recreate in 3D: an actual room, a landscape, an outdoor event. But for products, architectural models, or any scene you already have in Blender, synthetic capture wins on every metric.

• Lighting is fully controlled. No shadows from a passing car, no overexposed windows.
• Camera positions are mathematically precise. Synthra places cameras with known focal lengths, so training software reads them directly from transforms.json instead of estimating them from feature matching.
• No motion blur, no noise, as long as you use enough Cycles samples.
• You can fill coverage gaps instantly by adding manual cameras and re-rendering just those frames.

What Synthra generates

After a render, the output folder contains three things that every GS trainer needs:

• images/: all rendered frames as PNG files (0000.png, 0001.png, ...)
• transforms.json: camera positions, focal lengths, and frame paths in the format PostShot, Nerfstudio, and most GS trainers expect
• points3d.ply: a sparse point cloud in COLMAP format, exported from your Blender geometry, so the trainer has an instant reconstruction starting point without running COLMAP

Step 1: Install Synthra

Download the Synthra addon from your Luxa dashboard or from superhivemarket.com/products/synthra. In Blender go to Edit → Preferences → Add-ons → Install, select the .zip file, and enable it. The Synthra panel appears in the N-panel (press N in the 3D viewport).

Step 2: Choose the right camera mode

Synthra has five camera modes. The right one depends on what you're capturing:

• Object Capture: for a single product or object. Cameras orbit the subject from multiple elevations and distances. Best for chairs, shoes, jewelry, any standalone object.
• Indoor Volume: for rooms and interiors. Cameras fill a volume evenly with collision detection to avoid clipping into walls or furniture. Use this for architectural walkthroughs or showroom captures.
• Walkthrough Path: for linear sequences: corridors, streets, a path through a garden. Cameras follow a spline with consistent spacing.
• Hybrid: combines Path and Indoor in one capture. Half the cameras follow the path, the other half fill the surrounding volume. Recommended for complex interiors where you want both coverage and a clear walking route.
• Manual: place individual cameras by hand for specific angles the automatic modes miss.

Step 3: Render settings for Gaussian Splatting

This is where most people make mistakes. Two settings matter above everything else:

• Cycles samples: use at least 128. Too few and the image noise corrupts GS training: you get floaters and blurry regions. For a final dataset use 256–512. For framing checks only, use Fast Render ⚡ (20 samples), but never submit a Fast Render dataset to training.
• Denoising: turn it ON for GS targets (PostShot, Nerfstudio). Turn it OFF if you're using RealityCapture for SfM photogrammetry.

Step 4: Run the render

In Step 4 of the Synthra panel, enable "Camera data (transforms.json)". This is the file that ties every image to its camera position in space. Without it, your images folder is just a pile of pictures with no spatial meaning.

Click Start Rendering. Synthra renders all frames automatically and writes the output folder. When it finishes, click "Export points3d.ply" in the success panel to generate the point cloud from your Blender geometry.

Step 5: Train with PostShot or Nerfstudio

PostShot is the fastest path from dataset to trained splat. Drag your entire output folder onto the PostShot window: it detects transforms.json and points3d.ply automatically, no manual import needed. Training typically takes 5–15 minutes.

For Nerfstudio, point it at the same folder:

ns-train splatfacto --data /path/to/output_folder/

Step 6: Upload to Luxa

Once training is done, export the result as a .ply or .splat file. Open your Luxa dashboard, go to the Walk-in Gaussian Editor, and upload the file. In seconds you can walk through your scene in the browser at 60fps. From there, grab the embed code and put it on any website, or share the direct link.