The paper presents a state‑of‑the‑art method for line art vectorization that moves beyond heuristic stroke extraction. Tianhao Chen, Clara Fernandez, and Marteinn Oskarsson and their co‑authors leverage depth prediction and semantic feature extraction models to split a sketch’s skeleton graph into meaningful subgraphs, initializing strokes that better respect artistic intent. Strokes are modeled as Bézier curves with both geometric and appearance components, and 2D Gaussian splatting provides fast, differentiable rendering. This formulation enables efficient fitting of strokes to raster input while jointly optimizing control points and brush texture.

The approach also incorporates temporal tracking and adaptive keyframe insertion for video applications, delivering high‑quality reconstructions at speed. Users can easily correct splines and adjust optimization parameters, making the system highly interactive. The work was presented at SIGGRAPH (2026) and is authored by researchers from Disney Research|Studios, Walt Disney Animation Studios, and ETH Zurich.