Look Both Ways Before You Cross: Lifting Cross Fields From 2D Visual Priors

A new technique called CrossLift computes cross fields on 3D meshes by extracting directional guidance from 2D images, according to research posted to arXiv on 25 May 2026 [1]. The method uses text-to-image generative models to produce feature-aligned quad mesh renderings, then lifts the resulting per-pixel directions back onto the mesh surface [2]. The approach, detailed in a paper titled "Look Both Ways Before You Cross: Lifting Cross Fields From 2D Visual Priors," addresses a persistent challenge in computer graphics: generating surface-aligned directional fields that respect semantic features of an object [1]. CrossLift first synthesizes 2D images of quad meshes aligned with visual features using text-to-image priors [2]. It then extracts explicit per-pixel directions from those images and back-projects them onto the 3D mesh [1]. The team proposes custom confidence-based weights during two stages of smooth interpolation on the mesh surface — first within each view, then across multiple views — to resolve conflicts between candidate directions on the same face and to extend the field to occluded regions [2]. The method is modular and can operate with various 2D visual priors [1]. Beyond cross-field computation, the researchers demonstrate applications in texture-aligned quad meshing and interactive cross-field design using coarse, user-drawn lines as input signal [2]. Tests on a diverse set of organic and mechanical shapes produced quad meshes with superior semantic alignment compared to existing methods [1]. The project page is hosted at crosslift.github.io [2]. The work appears solely as a preprint on arXiv and has not been peer-reviewed [1].