BSD-GAN: Branched Generative Adversarial Network for Scale-Disentangled Representation Learning and Image Synthesis

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Generative Adversarial Networks (GANs) are powerful tools for image synthesis.
• Existing GANs often struggle to effectively learn and represent image features across multiple scales.
• Unconditional GANs lack explicit control over feature representation, limiting their application in complex generation and editing tasks.

Purpose of the Study:

• To introduce BSD-GAN, a novel multi-branch and scale-disentangled training method for unconditional GANs.
• To enable GANs to learn image representations at multiple scales for improved generation and editing.
• To demonstrate the ability to manipulate latent codes corresponding to different feature scales.

Main Methods:

• BSD-GAN employs a multi-branch training strategy with progressively increasing image resolutions.
• Input noise vectors are split into sub-vectors, each trained to learn representations at a specific scale.
• Sub-vectors are progressively unfrozen during training as higher resolutions and more network layers are introduced.

Main Results:

• BSD-GAN effectively learns scale-disentangled image representations without additional labels.
• The method synthesizes novel, high-resolution image content without compromising quality.
• Experiments validate the effectiveness of BSD-GAN for various image generation and manipulation applications.

Conclusions:

• BSD-GAN offers a robust method for multi-scale feature learning in unconditional GANs.
• The scale-disentangled representations facilitate direct manipulation of latent codes for enhanced control.
• This approach significantly advances the capabilities of GANs in image synthesis and editing.