🖱️ To interact with the 3D models, Scroll to zoom, hold Left Click + drag to rotate, and hold Right Click + drag to translate.
🔍 To zoom in on any visualization, Click the image or model to open it in a popup.
Method Overview. Taking a mesh as input, our method renders depth from multiple views using these to condition a 2D generative model which produces multi-view images depicting a quad pattern on the mesh. We extract the alignment directions from these images using pixel gradients and project those directions back to the 3D surface. Here, we interpolate using two solves, first per-view to accumulate the directions into a cross field over the faces and second to interpolate these multi-view cross fields into a single smooth cross field for the mesh.
Gallery. Our method produces desirable quads for a variety of shapes, both organic and manufactured. The quad lines do not just follow geometric features: our method generates quads whose edges follow semantic feature lines of the shapes through our use of text-to-image priors and without any 3D interaction.
Texture-Aligned Quads. Given a textured mesh (left), we show the 2D prior output (gray) and final quad mesh (blue) using either untextured (middle) or textured (right) renders of the mesh to condition our 2D prior. Using untextured renders produces quads aligned to the features of the shape, while using textured renders allow our method to align quads to the features in the texture.
Hand-Drawn Coarse 2D Guidance. Our method can extract signals from coarse hand drawn alignment directions on renders of our mesh. This 2D guidance is easy to create and provides fine-grained control to the user without requiring any 3D technical knowledge or skills.
Compatibility with Alternative 2D Guidance Signals. Our method is modular and while we use Flux (left) for the results shown in the main paper, we also show our method with alternative 2D priors Gemini 3 (middle) and ChatGPT 5.2 (right). Our method is robust to alternative 2D guidance signals (shown in gray) producing plausible semantically aligned quads in each case.
Robustness to Surface Noise. Our visually-guided method is robust to surface noise such as the wrinkles in the tablecloth. A baseline method using principal curvature constraints alongside a smoothness term aligns to these local features producing undesirable edge flow and introducing additional singularities.
Interpolation to Non-visible Mesh Regions. Our method robustly produces a smooth field even in regions that are not visible in the 2D views ($\ie$ the area under the penguin's wing). Without the smoothness term ($\lambda_s = 0$), the resulting field has holes and our quad mesh extraction fails. With this term ($\lambda_s = 1$), we are able to smoothly interpolate the surrounding field to get meaningful values in the unseen region.
✨ Hover over a streamline visualization to reveal the underlying mesh.
Streamlines. We visualize streamlines of our cross field traced across the mesh surface. These streamlines show the field alignment used to produce our final quad meshes.
Qualitative Comparison. We compare our method to QuadriFlow, QuadWild and NeurCross. QuadriFlow and QuadWild often produce wavy edge flow, while NeurCross tends to produce overly axis-aligned edge flow that fails to align to specific features. In contrast, our method produces edge flow that demonstrates superior alignment to both semantic and geometric features.
