MatLoom: Layered Text-to-Material Generation in a Compact Program Space
MatLoom turns text prompts into compact executable PBR material programs using pretrained language models, beating diffusion baselines.
Researchers introduce MatLoom, a compact layer-oriented language that turns text prompts into physically based material maps with pretrained language models and no task-specific fine-tuning. Parser-guided repair, preview critique, and noise-seed search revise designs while keeping source parameters editable. On 141 prompts and six backbones, the best configuration beat three diffusion baselines on four flat-layout alignment metrics, and a 30-participant blind study chose its renders 59.2% of the time versus 19.3% for the top baseline.
- Programs compose alpha-masked layers into PBR maps without task-specific fine-tuning
- Median retained program length is 21 lines across six backbones
- Best setup beat three diffusion baselines on all four alignment metrics
- Blind study of 30 people preferred its renders 59.2% of the time
From the original article
1 min read at arxiv.orgMaterial generation should produce not only an appearance, but also the rules that construct it. We introduce MatLoom, a compact, layer-oriented language for text-to-material generation with pretrained language models.
By Anson Y. Lam · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.