Turn Any 3D Model Into an Animated Pixel-Art Diorama With a Free Browser Tool
Hand-drawn pixel art has one stubborn limit: the camera does not move. A pixel artist draws a scene from one angle, and if you want another angle, someone draws it again, dot by dot.
Over the weekend of November 15, that rule started to look optional. Several free projects appeared on GitHub within about a day of each other, all making the same thing: tiny 3D worlds you can spin with your mouse that still look hand-placed, with a limited palette, crunchy dithering and hard outlines. Two of them credit the same source, posts by an artist who goes by Silicon Jungle. The recipe was simple enough to describe in a sentence, and within hours people had built it from scratch, each in a different way.
The version worth your time if you make things rather than code things is the one that takes a 3D model you already have, a scan, a sculpt, a game character or something an AI generator spat out, and converts it in one command. My position: a convincing pixel-art 3D look for your own models is now free, local and yours to change, and the cost is half an hour in a terminal.
What the look actually is
The style comes from stacking three old ideas.
Shapes. You start with 3D forms, either described with math (a sphere here, a tapered cone there) or taken from an ordinary 3D model file.
Splats. Instead of rendering those shapes as smooth polygon surfaces, the tools cover them in tens of thousands of tiny flat discs, like sequins sewn onto a costume. This is Gaussian splatting, the same idea behind the 3D scans phone apps now make. Each disc knows which way its patch of surface faces, so it can be lit live as the camera moves.
A pixel pass. The whole scene is drawn at a very low resolution, around 200 pixels tall, with the soft edges switched off. Every pixel then snaps to a small palette, around 15 colors in the default scenes, and smooth shading turns into a checkerboard dither pattern, exactly the trick 16-bit games used to fake gradients.
The result reads as pixel art because the pixel grid is fixed to your screen. Zoom in and the pixels do not get bigger; the world just gets closer. Orbit around a character and it stays crisp from every side.
Three ways in, one for each kind of creator
The weekend produced three usable entry points, and they suit different people.
If you have models: pixelart-gaussiansplat by Mert Cobanov converts any glTF or GLB file, the common 3D model formats, into a pixel-art diorama. If the model has a skeleton and animations, they keep playing. It can also build a little floating island under your model. MIT license, a permissive license that allows commercial work.
If you would rather ask than build: pixel-splats from the game platform Patch ships five sample worlds in a single web page and a skill (a plug-in for Claude's coding assistant, Claude Code) that builds new ones from a sentence. It can also film a scene straight to video. Apache 2.0 license.
If you like writing scenes by hand: SyntaxSplat lets you describe scenes in code, bakes the lighting in, and exports a standard splat file (.ply, the format 3D scan viewers open). MIT license.
There is also a fourth, from Silicon Jungle's own account, but it carries no license, which means you can look and learn but not build on it.
This article walks through the first, because it turns the work you already have into something new.
Put this into practice: your model as a pixel-art diorama
You need a computer with a recent browser, Node.js (a free program that runs JavaScript tools) installed, version 20.19 or newer, or 22.12 or newer if you are on the 22 line, and a 3D model in GLB or glTF format. Blender exports GLB directly, and so do most AI image-to-3D tools.
1. Install and open the viewer.
git clone https://github.com/cobanov/pixelart-gaussiansplat
cd pixelart-gaussiansplat
npm install
npm run dev
Open the address it prints. Drag to orbit, scroll to zoom, and try the built-in examples (a shrine, a lighthouse, a caravan camp, a toad in a swamp) to get a feel for the look.
2. Convert your model.
npm run convert path/to/model.glb mymodel
This samples the surface of your model and its colors into about 110,000 discs and adds it to the viewer's menu. Open /?m=mymodel in the browser.
3. Give it somewhere to stand.
npm run island mymodel ruins
The other themes are desert, swamp, crag and dais. The tool plays every animation on your model first, measures how far it reaches and where its feet land, then builds the island around it: flat where it stands, props where it does not reach. It takes ten to twenty-five seconds.
4. Tune the look in the address bar. Add options to the end of the web address: px=4 sets the pixel size, colors=24 the palette size, bands=5 the number of flat lighting steps (0 for smooth shading), light=1.3 brightens dark textures, anim=run starts a specific animation. If your model has several animations, keys 1 to 9 switch between them.
5. Record it. The record button in the panel saves a video of the canvas. For frame-perfect output, open the page with &capture=1 added: the clock stops, and the viewer will draw any exact moment you ask for. Use an even pixel size like 4 or 6 for video, because odd sizes blur at the edges once the video is compressed.
That last detail tells you something about the project. It was built by someone who actually exported clips and noticed the problem.
No model? Ask for a scene instead
If you do not have a 3D file to hand, the pixel-splats project gets you to the same look from a sentence. Its sample worlds are a treasure chest with teeth, a fox by a campfire, a split geode full of crystals, a lighthouse sweeping the sea and a koi pond. Each is a few dozen simple shapes; the fox is built from fourteen.
If you use Claude Code, install the project's skill with two commands, /plugin marketplace add patch-gg/pixel-splats and then /plugin install pixel-splats@pixel-splats, and ask for what you want. The README's own example is "Make a diorama of a witch's hut in a swamp, with fireflies and a bubbling cauldron." The skill works the way an illustrator does: build the scene, render it from three angles, look, fix the worst thing, repeat.
When you like it, node tools/film.mjs lighthouse --seconds 8 renders an 8-second clip at 1920 by 1080, or a GIF if you ask for one. The filming tools need Node 22, Chrome and the free video tool ffmpeg.
The trade is control. Motion here is limited to rigid moves plus four built-in wobbles (sway, water, swim and breathe), so you get a living diorama, not a character who acts.
Why this is worth an afternoon
It turns assets you already paid for into a second style for free. A character you modeled for one project, a product scan, a creature from an image-to-3D generator: each becomes a retro asset with real camera moves, which no drawn sprite sheet can give you.
It is consistent in a way AI image tools are not. Ask an image model for the same character from eight angles and you get eight slightly different characters. Here, every angle is the same model, so the design cannot drift.
It is yours. Everything runs in your browser on your machine. Change the palette, copy and modify the code, ship the result in a game or a music video. The MIT license allows it.
Where it breaks
It only reads color. The converter uses your model's base colors and lighting from its shape. It ignores the extra texture maps that make surfaces look bumpy or metallic, and it skips glass entirely. Shiny or detailed materials come out flatter than you expect.
More detail looks worse. The defaults are deliberately low. The docs explain that at normal pixel sizes, extra discs or sharper textures just make neighboring pixels disagree, and the surface looks sandy. Resist the urge to crank it.
Animation has limits. A character skeleton can have at most 256 bones, bones should not change size during an animation, and morph shapes (the sculpted smiles and blinks many characters use) are ignored. A character who emotes with morph shapes will lose those expressions.
The example creatures are not yours. The minotaur, toad, vulture and hyena in the repo come from a game called Kervan and are included as examples only. They are not covered by the MIT license, so do not put them in your work.
It is a few days old. One person maintains it, few people have bookmarked it on GitHub yet, and it has no track record. Treat it as a sketchbook until you have tested it on your own model and your own deadline.
It is still a terminal. There is no app. If installing Node and pasting commands sounds miserable, start with the pixel-splats demo on Patch instead and come back later.
A style you can own
What strikes me most about this weekend is not the look. It is how fast a look stopped belonging to one person's feed and became something anyone can run. A few posts described a recipe, and within a day there were several free machines that make it, each opening a different door.
Pick one model from your drive, something you made and like, and run it through the converter tonight. Turn it slowly in the viewer and see whether it suggests a project you would not have thought of yesterday. If it does, you have a new style. If it does not, you have spent twenty minutes and learned exactly what this look is good for.