Stable-Hair V2: Real-World Hair Transfer via Multiple-View Diffusion Model

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Diffusion models excel at generating diverse hairstyles but struggle with multi-view consistency.
• Generating consistent multi-view hair is crucial for applications like digital humans and virtual avatars.
• Existing methods lack robust solutions for high-fidelity, view-consistent hair transfer.

Purpose of the Study:

• To propose Stable-Hair v2, a novel diffusion-based framework for multi-view hair transfer.
• To achieve robust, high-fidelity, and view-consistent hair transfer across multiple perspectives.
• To establish a new benchmark for multi-view hair transfer.

Main Methods:

• Developed a comprehensive multi-view training data generation pipeline for triplet data.
• Integrated polar-azimuth embeddings for pose conditioning and temporal attention layers for view consistency.
• Designed a novel multi-stage training strategy including Pose-Controllable Latent IdentityNet, Hair Extractor, and Temporal Attention training.

Main Results:

• Stable-Hair v2 accurately transfers detailed and realistic hairstyles to source subjects.
• The method achieves seamless and consistent hair results across multiple views.
• Demonstrated significant outperformance over existing methods in extensive experiments.

Conclusions:

• Stable-Hair v2 establishes a new benchmark for multi-view hair transfer.
• The framework provides a robust solution for high-fidelity, view-consistent hair generation.
• The approach is highly suitable for digital human and virtual avatar applications.