Segmentation-Based Parametric Painting

We present a novel image-to-painting method enabling the generation of large-scale, high-fidelity paintings with varied human-like styles. Our approach leverages a segmentation-based painting process combined with a dynamic attention maps system, inspired by human painting techniques, to optimize brush strokes across different image regions. This allows for capturing both overarching structures and minute details, offering stylistic flexibility. Our method's efficiency in handling large canvases sets it apart, achieving superior aesthetic and functional results, validated through comprehensive evaluations. Code available at: https://github.com/manuelladron/semantic_based_painting.git