Efficient Style-Corpus Constrained Learning for Photorealistic Style Transfer.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Photorealistic style transfer is challenging, often producing unrealistic artifacts.
• Current methods incur significant computational expenses.

Purpose of the Study:

• To develop a novel scheme for photorealistic style transfer that enhances realism and reduces computational cost.
• To improve style consistency and minimize artifacts in stylized images.

Main Methods:

• Proposed a Style-Corpus Constrained Learning (SCCL) scheme utilizing style-specific and style-agnostic characteristics.
• Employed adversarial distillation learning to train a fast, simple network, replacing complex feature transform models.
• Introduced a style-corpus to ensure style consistency across different samples.

Main Results:

• The SCCL method generates rich-detailed, photorealistic stylized images.
• Achieved significant reductions in computational cost, with the method being 13–50 times faster than WCT².
• Demonstrated improved photorealism and style consistency compared to existing methods.

Conclusions:

• The proposed SCCL scheme effectively addresses the challenges of photorealistic style transfer.
• The method offers a computationally efficient and high-quality solution for image stylization.
• This work advances the state-of-the-art in generating realistic and consistent stylized images.