Controllable Sketch-to-Image Translation for Robust Face Synthesis

Area of Science:

• Computer Vision
• Image Synthesis
• Artificial Intelligence

Background:

• Synthesizing realistic images from sketches is challenging.
• Existing methods often lack fine-grained user control and robustness.

Purpose of the Study:

• To develop a controllable sketch-to-image translation framework.
• To enable interactive synthesis and editing of face images using hand-drawn sketches.
• To enhance realism while maintaining structural consistency with input sketches.

Main Methods:

• A novel dilation-based sketch refinement method inspired by coarse-to-fine painting.
• Multi-level refinement with a control parameter for sketch reliability.
• Scale-aware style transfer for modeling sketch features at different levels.
• Advanced user controls for editing regions, facial attributes, and non-uniform refinement.

Main Results:

• The framework successfully synthesizes and edits face images with high visual quality.
• Demonstrated robust user controllability for fine-grained semantic editing.
• Achieved a balance between output realism and structural consistency with input sketches.

Conclusions:

• The proposed method offers a novel and effective approach to controllable sketch-to-image translation.
• It provides significant improvements in user control and editing flexibility.
• The framework has potential for various applications in image synthesis and editing.