LineDrawer: Stroke-Level Process Reconstruction of Complex Line Art Based on Human Perception

Abstract

Line art is a fundamental yet powerful form of artistic expression. In this paper, we introduce a novel task aimed at enhancing novice understanding of reproducibility in line drawings: reconstructing the stroke-by-stroke drawing process from complex line art. This task poses substantial challenges, as it requires resolving stroke ambiguity, variations in stroke thickness, and stroke overlapping. To address these issues, we propose a hierarchical framework that emulates human drawing behavior, comprising three stages: (1) high-level generation of global semantic stroke order, (2) mid-level optimization of human drawing mechanics, and (3) low-level perceptual stroke rendering. Drawing inspiration from the human tendency to conceptualize the overall structure before refining local details, we first extract keyframes of the drawing sequence that guide global ordering using a diffusion-based model. Simultaneously, based on the assumption that humans can infer strokes from any cue point in a line drawing, we train a stroke renderer to extract variable-width sub-strokes at the pixel level. Lastly, we formulate a set of equations to model human drawing dynamics, enabling more detailed inference of stroke composition and sequencing within the identified keyframes. This framework effectively integrates high-level semantic understanding with low-level stroke reconstruction, facilitating stroke-level process recovery in complex line drawings. Extensive experiments and user studies demonstrate that our method produces relatively natural and coherent drawing process animations for high-quality line art.