DreamUV: Flow Matching Model Generates Artist-Quality UV Layouts for 3D Models

Researchers have introduced DreamUV, an end-to-end generative framework that reframes UV unwrapping as a flow matching problem. Rather than optimizing traditional geometric energy functions, DreamUV learns a mesh-conditioned transport process that maps noise samples to distributions of artist-like UV layouts — capturing structural patterns such as straightened seams and axis-aligned islands that professional artists prefer but that are difficult to encode in classical methods.

The system incorporates a boundary-aware training strategy that prioritizes seam geometry, alongside a Model-in-the-Loop Finetuning (MITL) scheme to stabilize sampling under heterogeneous supervision. Evaluated on a large-scale dataset of professionally authored UV layouts, DreamUV outperformed both classical and learning-based baselines on boundary straightness and island alignment, while maintaining competitive distortion metrics. A user study with professional artists confirmed the outputs align with practical production requirements.