Bézier Spline-Based Line Art Vectorization Using 2D Gaussian Splatting

We propose a principled approach that utilizes depth prediction and semantic feature extraction models to extract strokes aligned with artistic intent from raster images, partitioning the sketch's skeleton graph into subgraphs.

Strokes are modeled as Bézier curves with geometric and appearance components, employing 2D Gaussian Splatting for fast and differentiable rendering. This method enables the joint optimization of control points and brush textures, allowing for efficient fitting of strokes to the input image.

Key achievements include:

• Reconstruction Performance: Achieving SOTA (State-of-the-Art) results in terms of image reconstruction
• Efficiency and Quality: Fast processing speed and high-quality stroke generation
• User Control: Maintaining high compatibility with user input through spline editing and optimization parameter selection
• Video Scalability: Demonstrating promising results in video through the introduction of temporal tracking and adaptive keyframes