If you want to turn a photo into something you can spin around in 3D, there are two techniques worth knowing: photogrammetry and gaussian splatting. They solve the same problem in very different ways.
Photogrammetry
Photogrammetry reconstructs 3D geometry from 2D images by matching features across many overlapping photos. Move your camera around an object, and the algorithm triangulates where each point sits in space, then wraps a texture over the resulting mesh.
The output is a classic textured 3D model — vertices, faces, and a texture map. It is great for objects you can walk all the way around, but it needs coverage: gaps in your photos become holes in the model. Shiny, transparent, or featureless surfaces are hard because there's nothing consistent to match.
Gaussian Splatting
Gaussian splatting represents a scene as millions of tiny, soft 3D blobs ("splats"), each with a position, color, and transparency. Instead of building a mesh, it renders those blobs directly. The result looks remarkably photoreal and handles soft edges, reflections, and fine texture that meshes struggle with.
Because the representation is so flexible, a splat can be reconstructed from far fewer inputs — and with the right model, even from a single photo. That's the trick behind turning one image into a 3D scene.
Where Mukbang Uses Each
Mukbang uses both, depending on what you give it:
- Short video → photogrammetry. Circle your food for a few seconds and Mukbang reconstructs a full model from the motion. No LiDAR required.
- Single photo → gaussian splatting. One picture becomes a draggable 3D scene, rendered right in the browser by the <image-3d> web component.
Either way, you don't need a depth sensor, a scanning rig, or desktop software — just a phone camera.
