RPD-GAN: Learning to Draw Realistic Paintings with Generative Adversarial Network

Area of Science:

• Computer Vision
• Image Synthesis
• Artificial Intelligence

Background:

• Painting style transfer typically focuses on artistic styles, with existing methods like Neural Style Transfer (NST) and unsupervised cross-domain image translation often insufficient for realistic painting generation.
• Generating realistic paintings requires precise style capture while preserving original content features and image structures, a challenge not adequately addressed by current approaches.

Purpose of the Study:

• To develop a novel framework for automatic realistic painting style transfer that overcomes the limitations of existing methods.
• To generate realistic paintings, such as gouache, sketches, or pen-and-ink portraits, with enhanced content preservation and style accuracy.

Main Methods:

• Proposes Realistic Painting Drawing Generative Adversarial Network (RPD-GAN), an unsupervised cross-domain image translation framework.
• Decomposes image stylization into four stages: feature encoding, de-stylization, re-stylization, and decoding.
• Enhances the CycleGAN architecture by embedding content-consistency and style-alignment constraints in the feature space.

Main Results:

• RPD-GAN demonstrates superior performance in generating realistic paintings compared to existing methods.
• The embedded constraints significantly improve both content preservation and style-capturing capabilities.
• Extensive experiments validate the effectiveness and superiority of the proposed RPD-GAN framework.

Conclusions:

• RPD-GAN offers an effective solution for realistic painting style transfer, achieving high-quality stylization.
• The method successfully balances style accuracy with content preservation, advancing the field of image synthesis.