A new AI system called MatLoom generates textures as short programs you can edit, not just finished images.
Researchers built MatLoom, a compact programming language for producing physically based rendering material maps - the color, roughness, and bump channels used in 3D rendering - from text prompts. Rather than training a new model, it has an existing pretrained language model write programs built from alpha-masked layers, with each layer defining how pattern, color, and surface relief interact. A separate interpreter turns those programs into the actual material maps, and the code stays readable afterward, so an artist can open it and tweak named parameters by hand. The system also critiques its own preview renders and searches random seeds to refine results, all without any task-specific fine-tuning.
Tested on 141 prompts across six different language model backbones, MatLoom's best setup beat three diffusion baselines on every prompt-alignment metric the researchers measured, and in a blind test with 30 people it was picked 59.2% of the time versus 19.3% for the strongest rival. The real pitch isn't just better-looking textures - it's that each output is a roughly 21-line program an artist can open and modify, instead of a locked grid of pixels.
That tradeoff, generation as code instead of diffusion, is still only proven on a 141-prompt benchmark judged in part by the team that built the tool.