MD-Face: MoE-Enhanced Label-Free Disentangled Representation for Interactive Facial Attribute Editing

2026-04-22

AI Summary

This work addresses the challenges of attribute entanglement and high annotation costs in unsupervised facial attribute editing by proposing the first mixture-of-experts (MoE)-based disentangled representation learning framework. The method employs a dynamic gating mechanism to assign specialized experts, enabling the learning of semantically independent latent vectors. To enhance editability, it introduces a geometry-aware loss to align semantic boundary vectors (SBVs) and incorporates a Jacobian push-forward technique to improve control precision. Evaluated on ProGAN and StyleGAN, the approach significantly outperforms existing unsupervised baselines and achieves performance comparable to supervised methods. Moreover, compared to diffusion-based models, it delivers higher image fidelity with substantially lower inference latency, making it suitable for high-quality, real-time interactive editing.

Technology Category

• Machine Learning: Mixture of Experts (MoE)
• Computer Vision: Diffusion Models for Vision
• Search and Optimization: Mixed Discrete/Continuous Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications

Abstract

GAN-based facial attribute editing is widely used in virtual avatars and social media but often suffers from attribute entanglement, where modifying one face attribute unintentionally alters others. While supervised disentangled representation learning can address this, it relies heavily on labeled data, incurring high annotation costs. To address these challenges, we propose MD-Face, a label-free disentangled representation learning framework based on Mixture of Experts (MoE). MD-Face utilizes a MoE backbone with a gating mechanism that dynamically allocates experts, enabling the model to learn semantic vectors with greater independence. To further enhance attribute entanglement, we introduce a geometry-aware loss, which aligns each semantic vector with its corresponding Semantic Boundary Vector (SBV) through a Jacobian-based pushforward method. Experiments with ProGAN and StyleGAN show that MD-Face outperforms unsupervised baselines and competes with supervised ones. Compared to diffusion-based methods, it offers better image quality and lower inference latency, making it ideal for interactive editing.

Problem

• attribute entanglement
• disentangled representation
• label-free learning
• facial attribute editing
• annotation cost

Innovation

• Mixture of Experts
• disentangled representation
• label-free learning
• geometry-aware loss
• facial attribute editing