SwiftScene: External Importance Guidance for Hierarchical Visual Abstraction in Vector Sketch Generation

Sammanfattning

Vector sketch generation remains comparatively underexplored, especially for complete scene-level sketches. Existing methods can produce visually convincing outputs, but they provide limited explicit control over where visual detail is placed. As a result, the distribution of detail is typically determined by implicit model priors or fixed design choices, rather than by an externally specified notion of importance. This thesis presents SwiftScene, a modular framework for importance-aware scene-level vector sketch generation. Given an input image and an importance field, the framework decomposes the scene into regions, assigns stroke budgets according to regional importance, and generates per-region vector sketches using a fast object-level sketching primitive. The resulting scene sketch is therefore intended to reflect the supplied importance signal through its spatial distribution of strokes. To make this form of controllability measurable, the thesis introduces β(ℐ, S), a Kullback-Leibler-based divergence metric comparing the target importance distribution with the sketch's regional stroke distribution. Controlled experiments show that β behaves as an interpretable alignment measure, that modifying the importance field redirects sketch detail without retraining the sketch generator, and that the reference implementation produces scene-level sketches at inference times practical for interactive use. Together, the framework and metric reframe scene-level vector sketch generation as an alignment problem: not only whether a system can sketch a scene, but whether the resulting sketch allocates detail according to a specified visual importance signal.