Weakly Supervised Exaggeration Transfer for Caricature Generation With Cross-Modal Knowledge Distillation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Art

Background:

• Caricature generation aims to create exaggerated, artistic portraits from photos.
• Existing methods struggle with diverse, controllable exaggeration and identity preservation.
• Unsatisfactory and uncontrollable results limit the application of previous techniques.

Purpose of the Study:

• To propose ETCari, a weakly supervised network for diverse exaggeration transfer in caricature generation.
• To enable learning of various artistic styles from different artists for personalized caricatures.
• To achieve diversified exaggeration while effectively retaining facial identity features.

Main Methods:

• Developed a novel weakly supervised exaggeration transfer network named ETCari.
• Utilized TPS (Thin Plate Spline) control point deformation field as ground truth for loss calculation.
• Employed cross-modal knowledge distillation from a teacher model to a student model for simplified inference.

Main Results:

• ETCari successfully learns diverse exaggeration styles from multiple artists.
• The method achieves diversified exaggeration effects while preserving identity features.
• Experiments on the WebCaricature dataset confirm the effectiveness of ETCari in generating appealing caricatures.

Conclusions:

• ETCari offers a robust solution for controllable and diverse caricature generation.
• The weakly supervised approach addresses the challenge of limited labeled data.
• The proposed method enhances customization and artistic quality in caricature creation.