Disney Research and ETH Zurich Debut AI-Driven Line Art Vectorization
A new pipeline combining 2D Gaussian splatting and Bézier curves converts hand-drawn sketches into high-fidelity, editable vector graphics.
Disney Research Studios and ETH Zurich have developed a state-of-the-art method for vectorizing line art and sketches, presented at SIGGRAPH 2026. The system replaces traditional heuristics with a principled pipeline that converts raster pixels into mathematical paths while preserving artistic intent.
The new approach utilizes depth prediction and semantic feature extraction to initialize strokes. By splitting skeleton graphs into subgraphs, the system can more accurately identify where individual strokes begin and end. These strokes are modeled as Bézier curves, and the researchers implemented 2D Gaussian splatting to enable fast, differentiable rendering. This allows the system to optimize both the control points of the curves and the brush textures simultaneously.
The Challenge of Vectorization
Vectorization—the process of converting raster images into vector graphics—has remained a persistent challenge in computer graphics. While simple shapes are easily handled, hand-drawn sketches are notoriously difficult to vectorize because the software often fails to recognize the original 'artistic intent' of the stroke. Traditional tools frequently produce jagged paths or an excessive number of control points, making the resulting vectors difficult for artists to edit manually.
Impact on Digital Art
By combining Bézier curves with 2D Gaussian splatting, the researchers have created a differentiable pipeline that allows for the joint optimization of geometry and appearance. This technical shift makes the vectorization process significantly more efficient and ensures the output is highly compatible with manual user corrections. For digital artists and animators, this could drastically streamline the pipeline from initial sketch to final production asset, reducing the time spent on tedious manual cleanup.
Future Applications
Beyond static images, the system has already demonstrated promising results on video. The researchers achieved this through the use of temporal tracking and the introduction of adaptive keyframes, suggesting the method could eventually be used to automate the vectorization of entire animation sequences. While the core method is now established, the industry will be watching to see how these tools are integrated into commercial creative software.