SFXGraph: Causal-Compositional Graph Learning for Stylized Sound Effects in Webtoons

Abstract

Webtoons, a globally popular form of digital comics, rely on stylized sound effects (SFX) as essential visual cues for actions, emotions, and atmosphere. Since Webtoon SFX tightly entangle linguistic content, geometry, style, and scene context, conventional optical character recognition (OCR) systems designed for regular text often fail on these highly stylized effects. We propose SFXGraph, a causal-compositional graph parsing framework that models each SFX instance as a structured graph of linguistic units, geometric deformation, visual style, and contextual interaction factors, rather than decoding it as a flat text sequence. To learn this representation, we introduce counterfactual factor consistency, which separates scene-invariant content from SFX-induced factors and enforces the predicted graph to be sufficient for SFX removal, reconstruction, and controlled synthesis. We further construct a graph-supervised synthetic SFX generation pipeline that provides dense node, edge, style, and deformation supervision at scale. Experiments on public and custom Webtoon datasets demonstrate that SFXGraph consistently outperforms conventional OCR engines and vision--language baselines, while providing interpretable graph-level representations and stronger robustness to diverse artistic styles.