Neural Rendering for Game Character Auto-Creation.

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Video games often feature character creation systems allowing players to customize facial appearance.
• Existing methods for 3D face reconstruction primarily focus on mesh generation and often neglect user interaction.
• Automatic generation of realistic and customizable game characters remains a challenge.

Purpose of the Study:

• To propose a novel method for automatic game character creation from a single face photograph.
• To enable fine-grained facial parameter control for user customization.
• To integrate with game engines for seamless character generation.

Main Methods:

• Framing character creation as a self-supervised learning problem.
• Utilizing differentiable neural rendering for optimization.
• Introducing an "imitator" network to approximate non-differentiable game engine rendering processes.
• Generating fine-grained facial parameters with physical significance.

Main Results:

• The method successfully generates highly robust and vivid game characters.
• Experimental results on large-scale face datasets demonstrate the method's effectiveness.
• The system has been successfully applied in two commercial games, serving over 10 million users.

Conclusions:

• The proposed method offers an effective solution for automatic game character generation from single images.
• It provides users with intuitive control over facial parameters for enhanced customization.
• The approach integrates seamlessly with game development pipelines and has demonstrated significant real-world application.