From Pixels to Physical Objects

Most people think of 3D Gaussian splatting as a rendering technique. You capture a scene, train a model, and get a real-time interactive view. That is where the story usually ends. But a growing number of artists and engineers are pushing further, asking a genuinely strange question: what if you could hold it?

Printing a Gaussian splat as a physical object is not science fiction anymore. It is an emerging workflow that sits at the intersection of neural rendering, voxel art, and industrial inkjet printing. The results look like insects trapped in amber, or tiny dioramas frozen in time. And the process is more accessible than most people realize.

This article walks through the full pipeline, from capturing a scene to placing a printed crystal on your desk. It also covers the tools and decisions that matter most along the way.

What Makes Gaussian Splats Hard to Print

Standard 3D printing works with solid meshes. You define surfaces, slice them into layers, and the printer deposits material. Gaussian splatting does not work that way. A splat model is a cloud of millions of tiny ellipsoids, each carrying color, opacity, and orientation data. There are no hard surfaces. There are no clean edges.

That creates two immediate problems. First, most slicers have no idea what to do with a splat file. Second, the color in a typical splat is view-dependent. Tilt your head and the color shifts. A physical print cannot do that. You need to bake in a fixed color before the print job even starts.

The solution most workflows use is voxelization. The splat cloud gets converted into a 3D grid of small cubes, similar to how Minecraft represents a world. Each voxel stores a color value and a transparency level. That grid is something a printer can actually work with.

The transparency piece is especially tricky. Fur, hair, and thin translucent surfaces look wrong when voxelized at low resolution. Fine detail clumps together. Getting that right requires careful tuning of both the training parameters and the voxelization resolution.

Training Your Splat for Printing, Not Just Viewing

If you are capturing a scene specifically for printing, train differently than you would for a viewer. A few decisions made during training will save hours of post-processing later.

Use Spherical Harmonics at Level 0

Spherical harmonics control how view-dependent the color of each splat is. Level 3 gives you rich, realistic color shifts as you move around a scene. Level 0 locks in a single diffuse color per splat, with no view dependence at all.

For printing, level 0 is almost always the right choice. You cannot print view-dependent color. Averaging it out after training produces muddier results than training with level 0 from the start.

You can set this in most Gaussian splatting frameworks. In the original 3DGS codebase, it is a single parameter change in the training config. Newer frameworks like Nerfstudio expose this through their CLI as well.

Linear vs. sRGB Color Space

Training in linear color space can produce more physically accurate transparency values. The math behind opacity blending works more naturally in linear space. The tradeoff is that printing pipelines often expect sRGB input, so you may introduce a conversion step that adds complexity.

If you are working with a third-party printing service, ask them what color space their pipeline expects before you start training. Mismatched color space assumptions are one of the most common causes of prints that look too dark or too saturated.

# Setting SH degree to 0 in a typical Gaussian splatting training call
python train.py \
  --source_path ./data/my_scene \
  --model_path ./output/my_scene \
  --sh_degree 0 \
  --iterations 30000

Researchers exploring neural rendering pipelines often publish their training configs on Papers With Code, which is a useful reference when you are tuning for a specific output target like physical printing.

The Hardware Behind Crystal Printing

Not every 3D printer can do this. Consumer FDM printers (the kind that melt plastic filament) cannot produce transparent multi-color voxel objects. You need an industrial PolyJet printer, which jets tiny droplets of photopolymer resin in multiple colors simultaneously, then cures them with UV light layer by layer.

A Gaussian splat point cloud being voxelized into a grid for 3D printing

The Stratasys J850 Prime is one of the machines capable of this kind of print. It supports a wide gamut of colors including transparent materials, which is what makes the amber-like crystal effect possible. These machines are not cheap or widely available, which is why most people work through a printing service rather than owning one.

3D Printing Gaussian Splat Process

Capture Scene

Use a model to capture the desired scene.

Train Model

Train a neural rendering model for visualization.

Voxelization

Convert splat cloud into a 3D grid of voxels.

Bake Colors

Set fixed colors for the print job.

Print Object

Use industrial inkjet printer to create the physical object.

The voxel data gets sent to the printer as a series of layered images. Each layer is essentially a 2D cross-section of the object, with each pixel encoding which combination of inks to jet. The printer works its way up from the base, building the object one thin layer at a time.

Tools and Services Worth Knowing About

The tooling ecosystem for splat-to-print workflows is still young. Here is a practical breakdown of what exists right now.

Crysta.ai: One of the first services to offer a direct Gaussian splat printing pipeline. They handle voxelization and printing, and have an alpha editor for positioning and scaling your splat before printing. Check out their work on X (@Crysta_AI).
MagicaVoxel: A free voxel editor that is widely used for creating voxel art. While it does not natively import splat files yet, it exports formats compatible with PolyJet slicers. Bidirectional splat support would make it a powerful tool in this pipeline.
Nerfstudio: An open-source framework for training neural radiance fields and Gaussian splat models. It supports exporting point clouds and has an active community building custom exporters.
Polycam and Luma AI: Capture apps that produce Gaussian splat outputs from phone video. Both are solid starting points if you do not want to run training locally.
PyTorch: Most Gaussian splatting training code runs on PyTorch. If you want to customize your training pipeline, the PyTorch tutorials are a solid foundation.

For teams building their own voxelization tools, Hugging Face hosts several models related to 3D scene understanding that could accelerate custom pipeline development.

What the Output Actually Looks Like

The printed object is a small block of clear resin with the captured scene embedded inside it. Light passes through the transparent material, making it glow when held up to a lamp. The effect genuinely resembles an insect preserved in amber, which is part of why people find it so compelling.

Color accuracy depends heavily on the training decisions described above. Prints trained with level 0 spherical harmonics and correct color space matching tend to be the most faithful. Fine transparent geometry, like insect fur or hair, remains the hardest detail to preserve. The voxel resolution simply cannot capture sub-millimeter translucency without significant data.

One practical workaround some developers use is to capture objects with a neutral background and strong, even lighting. This reduces noise in the splat and gives the voxelizer cleaner opacity data to work with. If you are photographing a small object for printing, treat it like product photography. Even lighting matters enormously.

Speaking of product photography, if you need to clean up reference images before training, an AI background remover can help you isolate your subject cleanly. Cleaner training images mean fewer splat artifacts in the final model.

Practical Workflow: Capture to Crystal

Here is a condensed end-to-end workflow for anyone who wants to try this.

Capture your subject. Use a turntable setup with consistent lighting. Shoot 60 to 150 overlapping photos from multiple angles. Avoid reflective surfaces.
Train with SH level 0. Use Nerfstudio or the original 3DGS codebase. Set sh_degree 0 and train in linear color space if your printing service requests it.
Export your splat. Most services accept .ply files. Check with your provider before exporting.
Upload to a printing service. Services like Crysta.ai handle voxelization and slicing. Use their editor to scale and position your object.
Review a voxel preview. If the service offers one, check for clumping in fine geometry before approving the print job.
Print and wait. Industrial PolyJet prints typically take hours to days depending on queue and size.

The total cost varies by object size and service, but small prints in the few-centimeter range are accessible for hobbyists. Larger or more detailed prints scale up quickly in both time and cost.

Where This Technology Is Heading

The current state of splat printing is roughly where desktop 3D printing was in the early 2010s: technically impressive, somewhat finicky, and clearly on a trajectory toward becoming more accessible. A few developments would accelerate that significantly.

Better voxelization algorithms that preserve fine transparency are the most important bottleneck right now. Research in neural scene representation is moving fast. Groups publishing at venues tracked by Google AI Research and similar labs are actively working on higher-fidelity volumetric representations that could translate directly to better print quality.

Challenges in Gaussian Splat Printing

✓

No Hard Surfaces
Splat models lack defined surfaces.

✓

View-Dependent Color
Colors shift with viewing angle.

✓

Slicer Compatibility
Most slicers cannot process splat files.

✓

Transparency Issues
Fine details may get clumped in voxelization.

Direct voxel import in tools like MagicaVoxel would also open the workflow to a much broader audience of voxel artists. Right now the pipeline requires a Gaussian splat as input. Allowing raw voxel data from any source would let illustrators and game artists create printable crystal objects without ever touching a neural network.

On the business side, more printing services will enter this space as the workflow matures. That competition will drive down prices and improve tooling. For now, it is worth watching early movers and experimenting with your own captures.

Sharing and Documenting Your Prints

Once you have a physical print, how do you share it? Video is the most effective medium. Rotating the object under different lighting conditions shows off the transparency and depth in ways that static photos cannot. A short clip of light passing through the crystal is genuinely striking.

Close-up of a PolyJet 3D printer depositing transparent resin layers

If you are posting video online, file size matters for upload speed and quality. A free video compressor can cut your clip down to a manageable size without visible quality loss, making it easier to share across platforms.

For still photos of your print, keeping file sizes lean is equally useful. Run your reference images through a free image compressor before uploading to a portfolio or social media. It keeps your pages fast and your storage costs low.

If you are building an audience around your work, a free social media audit tool can help you understand which platforms are actually driving engagement for your content. Niche technical work like this tends to find its audience in unexpected places.

Budgeting for a Print Project

Printing costs add up faster than most people expect. A small test print might run a few dozen dollars. A high-quality final print of a complex subject can run into hundreds. Factor in your time for capture, training, and iteration as well.

If you are tracking expenses across multiple creative projects, a free budget tracker makes it easy to see exactly where your money is going each month. The tool supports custom categories, so you can track hardware, software subscriptions, and printing costs separately without juggling spreadsheets.

For teams or studios exploring this as a product line, the cost structure is worth analyzing carefully before scaling. What does it cost per unit at different volumes? Where are the margins? Those questions matter before you invest in a production workflow.

The Real Appeal of Physical Neural Renders

There is something philosophically interesting about this whole process. A Gaussian splat is a mathematical approximation of light. It is not a photo and not a mesh. It is a learned representation of how photons behaved around an object at a specific moment in time.

Printing that representation creates an object that has never existed in physical form before. The bee you print was not sculpted. It was not modeled. It was inferred from photographs, compressed into a cloud of ellipsoids, then voxelized and jetted in resin. That is a genuinely new kind of artifact.

Whether that matters to you depends on why you are making things. For artists, it is a compelling new medium. For engineers, it is a hard technical problem with interesting subproblems. For everyone else, it produces objects that are simply beautiful to look at.

If you want to go deeper on the math behind neural rendering, the Hugging Face documentation and community spaces host a growing number of 3D-focused models and papers. It is a good place to find collaborators and reference implementations.

Get Started Today

The barrier to entry is lower than it looks. You need a camera, a subject, a GPU for training, and access to a printing service. All the software in the core pipeline is open source.

Start small. Capture a single object with good lighting. Train with level 0 spherical harmonics. Export a .ply file and upload it to Crysta.ai or a comparable service. Review the voxel preview carefully before approving the print. Iterate once or twice on training parameters if the first result is not what you hoped for.

The community around Gaussian splatting is active and helpful. Questions tagged on Stack Overflow under gaussian-splatting get answered quickly. The GraphQL community has also developed interesting tooling for 3D data APIs, and the GraphQL documentation is worth reading if you are building a service around splat data pipelines.

Physical neural renders are one of the genuinely new things happening in creative technology right now. Try it before it becomes mainstream.