DualG-GAN, a Dual-channel Generator based Generative Adversarial Network for text-to-face synthesis

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Synthesis

Background:

• Text-to-image synthesis has advanced significantly, but its application to face generation remains underexplored.
• Existing methods often produce low-quality faces with poor text consistency.

Purpose of the Study:

• To develop an improved text-to-face synthesis method addressing quality and consistency issues.
• To introduce a novel generative adversarial network for high-fidelity face generation from textual descriptions.

Main Methods:

• Proposed DualG-GAN, an end-to-end dual-channel generator based generative adversarial network.
• Introduced a dual-channel generator block and a novel loss function for improved semantic similarity.

Main Results:

• DualG-GAN achieved state-of-the-art results on the SCU-Text2face dataset.
• Demonstrated superior performance in Fréchet inception distance (FID) and R-precision metrics compared to existing methods.

Conclusions:

• DualG-GAN significantly improves text-to-face synthesis quality and text-image consistency.
• This work establishes a new baseline for future research in text-to-face generation.