From Pixels to Plastic: The Wild World of 3D Printing Gaussian Splats
For decades, 3D printing has been a slave to the mesh. If you didn't have a watertight shell of triangles, you didn't have a print. But the rise of Gaussian Splatting (3DGS) is flipping the script. We are witnessing a shift from traditional modeling toward physicalizing neural reconstructions—essentially turning "fuzzy" digital clouds into tangible objects. It’s a messy, experimental frontier where AI meets additive manufacturing.
The Translation Layer: From Ellipsoids to Resin
The big hurdle with Gaussian Splats is that they aren't solid. They are collections of semi-transparent ellipsoids designed for high-speed rendering, not for holding a shape in a resin tank. To bridge this gap, creators are getting creative with geometry. Developer Wyatt Roy recently made waves by approximating 3D Gaussian functions using 14-face isospheres. By post-processing these shapes into a printable mesh, he’s effectively created a way to "print a photo" in three dimensions.
Other pioneers, like Dany Bittel, are tweaking the training process itself. Because 3D printers can't handle view-dependent colors (the way a splat looks different from different angles), Bittel experimented with training splats using level 0 spherical harmonics. This flattens the color into something a printer can actually understand, while training in linear space helps achieve more physically plausible transparency for the final output.
Splatcubes and the Future of Neural Objects
We are also seeing the emergence of "Splatcubes." This process involves taking a photogrammetry scan, converting it into a 3DGS dataset, and then filtering those splats to be printed inside a solid resin block. It’s a high-tech evolution of the lithophane, providing a level of realism that traditional voxel or mesh-based scans struggle to match.
Platforms like SuperSplat and Spline are already making it easier to edit and export these splats, while specialized pipelines from companies like cysta.ai are emerging to improve print fidelity directly from raw training files. While researchers are still perfecting voxel-based reconstruction methods like "Poxel" to address the limitations of neural fields, the momentum is clearly behind making these ethereal digital captures physical.
As we move away from manual modeling and toward these "captured" geometries, the line between photography and fabrication is blurring. We aren't just printing models anymore; we’re printing the very light and volume of the real world.
Sources
Media
