Core Concepts
U-Sketch introduces a U-Net architecture for efficient sketch-to-image synthesis, improving realism and reducing denoising steps.
Abstract
U-Sketch addresses the challenge of sketch-to-image synthesis by introducing a U-Net latent edge predictor.
The framework efficiently captures spatial correlations and simplifies input sketches for enhanced outputs.
Experimental results show U-Sketch outperforms MLP, reducing denoising steps by 80% while maintaining realism.
User studies confirm U-Sketch's superiority in terms of realism, edge fidelity, and overall structural coherence.
The impact of sketch simplification and noise initialization on image quality is explored.
Future work includes developing better evaluation metrics and investigating noise effects further.
Stats
Employing a U-Net architecture for edge map prediction.
Reduction of denoising steps by 80%.
User preference for U-Sketch over MLP in realism, edge fidelity, and structural coherence.
Quotes
"Our proposed U-Sketch leads to more realistic results aligned with spatial outlines."
"The U-Net latent edge predictor improves image synthesis efficiency and quality."