Controllable Image Synthesis With Attribute-Decomposed GAN.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Controllable image synthesis requires methods to manipulate specific image attributes.
• Existing generative adversarial networks (GANs) often struggle with fine-grained control over multiple attributes simultaneously.

Purpose of the Study:

• To propose Attribute-Decomposed GAN (ADGAN) and ADGAN++ for realistic and controllable image synthesis.
• To enable flexible and continuous control over image attributes by embedding them as independent codes in the latent space.

Main Methods:

• ADGAN uses two encoding pathways to decompose attributes into latent codes for simultaneous processing.
• ADGAN++ employs a serial encoding strategy and residual blocks with segmentation-guided instance normalization for complex attribute synthesis.
• Component layouts are extracted via a semantic parser and processed by a global texture encoder.

Main Results:

• Both ADGAN and ADGAN++ demonstrate superior performance in pose transfer, face style transfer, and semantic image synthesis.
• ADGAN++ effectively handles a large number of attribute categories and reduces computational costs.
• The methods achieve disentanglement of attributes for flexible control and component attribute transfer.

Conclusions:

• ADGAN and ADGAN++ offer advanced solutions for controllable and realistic image synthesis.
• The proposed serial encoding strategy in ADGAN++ addresses limitations of simultaneous attribute processing for complex tasks.
• The methods provide state-of-the-art results and enable fine-grained attribute manipulation in generated images.