Autoregressive models have proven effective for triangular mesh generation by converting connectivity prediction into a sequence of next-face token predictions. Quad-only meshes, however, are more challenging because quad faces impose stricter topological constraints. The model must serialize quad faces into a valid token sequence and also support unambiguous connectivity reconstruction during de-serialization.
Q-MAGIC addresses this with a tailored serialization scheme that requires at least two new vertices for the next face, reflecting the less flexible wiring of quads. It further proposes a singularity-aware serialization that prioritizes wiring from singular points, leading to more coherent traversals and shorter sequences. Geometry is encoded via a circular relative offset tokenization that relies mainly on local relative offsets.
An autoregressive Transformer learns the underlying distribution of quad meshes through next-token prediction. In experiments, Q-MAGIC produces more compact sequences while consistently improving geometric fidelity and mesh quality compared to state-of-the-art methods, demonstrating the effectiveness of the proposed representation and serialization strategies.