Multi-View Face Synthesis via Progressive Face Flow.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Computer Graphics

Background:

• Generative Adversarial Networks (GANs) struggle with identity preservation and faithful texture reproduction in multi-view face synthesis, especially under large angle rotations.
• Existing methods often fail to maintain facial details and identity when generating faces from significantly different viewpoints.

Purpose of the Study:

• To develop a robust method for multi-view face synthesis that overcomes limitations in texture fidelity and identity preservation.
• To propose a novel approach that effectively handles large angle rotations in face synthesis.

Main Methods:

• A Face Flow-guided Generative Adversarial Network (FFlowGAN) is proposed, comprising a face flow module for dense correspondence and a synthesis module for texture emphasis.
• The method divides large-angle synthesis into a series of small-angle rotations, each guided by computed face flow to preserve facial details.
• FFlowGAN is trained end-to-end, cascading multiple small-angle synthesis steps to ensure progressive propagation of facial features and identity.

Main Results:

• The proposed divide-and-conquer strategy significantly improves the faithfulness of facial texture and identity preservation during large angle rotations.
• FFlowGAN demonstrates superior performance compared to state-of-the-art methods on four benchmark datasets.
• Qualitative and quantitative evaluations confirm the effectiveness of the face flow guidance in maintaining salient facial features.

Conclusions:

• The FFlowGAN method offers a significant advancement in multi-view face synthesis, particularly for challenging large angle rotations.
• Breaking down complex synthesis into guided small-angle steps is an effective strategy for maintaining identity and texture fidelity.
• The proposed approach provides a robust solution for generating realistic and consistent multi-view face images.