CMU Geometer Keenan Crane Grows a 3D Dragon from 27KB of LLM-Written GLSL

CMU geometrician Keenan Crane ran an unconventional 3D modeling experiment: while most people use LLMs alongside Blender/Three.js to generate models, he had an LLM (Astra) directly produce 27KB of GLSL code defining a closed analytic implicit surface (a signed distance function), ultimately creating "Inigo the Dragon," a tribute to Inigo Quilez. He makes clear this was just an experiment—the efficiency falls far short of diffusion mesh generators—but it demonstrates the unique value of the LLM procedural-modeling route.

Confirmed

• The full workflow is public: the author shared the initial prompt given to Astra, which asked for a procedural SDF representation of the reference images in the style of Inigo Quilez's distance functions, with the model required to build its own library of basic shapes and shape-combination operations
• Step one used Nano Banana to generate front/side reference images, which the author called the easiest part of the whole pipeline—another sign of diffusion models' strength at dense visual content generation
• Total cost: 25 hours of AI calls and 27 rounds of human feedback; Astra's first direct generation took 45 minutes 47 seconds, and the author described the result as "uh, beautiful"—a stark gap versus a diffusion model's roughly 3-minute output
• Visualizing implicit surfaces is extremely expensive: even with a custom Metal renderer written to keep the machine from crashing, rendering a video still took about 45 minutes; in the end he had to convert the surface to a mesh to make inspection and editing practical
• The model is named "Inigo the Dragon"; the author admits probably only iq (Inigo Quilez) could build such a model by hand, and the workflow draws heavily on his work

Why it matters

• The author analyzed two reasons procedural 3D models are hard for LLMs to learn: such data has an irregular tree-like structure that is harder to learn than SDFs sampled on regular grids, and training data is scarce—hardly any artists model this way
• He still thinks it's worth trying: LLMs are inherently good at capturing statistical regularities in tree-like data, and VLM 3D spatial intelligence is improving rapidly; this route amounts to a "Blender primitives" for smooth organic shapes
• He also points out that inferring 3D from 2D projections is inherently ill-posed—many combinations of geometry, materials, lighting, and cameras can produce the same image—so regularization or priors are needed; his geometric prior is that "objects are mixtures of basic geometric primitives such as boxes and ellipsoids"
• The author explicitly cautions: if you really need organic 3D models from images, diffusion mesh generators like Meshy or Tripo are faster, cheaper, higher quality, and fully automatic—meshes take about 3 minutes and cost roughly 50 cents each

2026-10-10 ~ 2026-10-10 · 13 related posts

• [source] This 3D dragon is 27KB of LLM-generated GLSL, not a mesh or NeRF — keenanisalive · 2026-10-10
• CMU prof uses LLMs to write 27KB of GLSL defining an implicit-surface 3D dragon — keenanisalive · 2026-10-10
• [source] 25 hours of AI calls, 27 rounds of feedback: a dragon honoring Inigo Quilez — keenanisalive · 2026-10-10
• Real image-to-3D? Meshy/Tripo beat the 27KB GLSL implicit surface by far — keenanisalive · 2026-10-10
• Rendering an implicit surface: 45 minutes for one clip even with custom Metal renderer — keenanisalive · 2026-10-10
• Step one of the implicit-surface dragon: Nano Banana reference images — keenanisalive · 2026-10-10
• Reconstructing 3D from 2D is ill-posed; Nano Banana made the reference views — keenanisalive · 2026-10-10
• Tree-like procedural 3D data is hard for LLMs to learn — and data is scarce — keenanisalive · 2026-10-10
• Why it's still worth trying: LLMs handle tree-like data and VLM spatial skills are rising — keenanisalive · 2026-10-10
• [source] Full prompt revealed: making an LLM build procedural SDF 3D models from one image — keenanisalive · 2026-10-10
• LLM took 45 minutes to model a dragon; a diffusion model did far better in 3 — keenanisalive · 2026-10-10
• Telling an LLM to "believe in yourself" helps it write 3D SDF models, but not enough — keenanisalive · 2026-10-10

1 near-duplicate retellings: keenanisalive