SDFSpace

Narration: Supertonic 3 neural TTS, voice F2, local ONNX.

00:00:00.000 | Start with any mesh
SDFSpace starts with any triangle mesh. Here it is the textured Inigo dragon, half a million triangles with eight thousand pixel textures.

00:00:09.375 | Turn it into a distance field
A sparse grid of small bricks stores the distance to the surface. It builds in about a minute, runs directly on the GPU, and stays within 0.03 percent of the surface.

00:00:20.833 | A prompt becomes code
To change the shape, a coding agent writes a few lines of GLSL that compose with the original field. Adding a horn is a capsule and a smooth union.

00:00:30.333 | The agent looks at its work
GPT renders previews on a coordinate grid, looks at them, and revises. Claude reviews the result independently and sends back what is wrong, until the request is met.

00:00:41.417 | Back to a textured mesh
The edited field is extracted on the GPU, unwrapped in parallel, and the original textures are projected onto the new surface. The renders match the source at 49 decibels.

00:00:53.667 | Real requests
8 of 11 requests pass the reviewer. A single blind pass averages 4.7 out of ten. The loop averages 7.7.

00:01:03.542 | Open source
Code, the benchmark suite and every edit transcript are open source. Try it at sdfspace dot app dot nz.