StyleAutoEncoder for Manipulating Image Attributes Using Pre-trained StyleGAN

2024-12-28

Pacific-Asia Conference on Knowledge Discovery and Data Mining

AI Summary

To address fine-grained style editing of pre-trained StyleGAN images, this work proposes a fine-tuning-free, unsupervised, single-shot-driven disentangled encoding framework. Methodologically, it achieves the first unsupervised linear separation and interpretable semantic mapping of attributes in the StyleGAN latent space; this is accomplished via feature distillation, contrastive latent-code clustering, an attribute-aware dual-path autoencoder, and gradient-guided optimization of semantic directions—enabling high-precision attribute localization and disentanglement. Evaluated on FFHQ and AFHQ, the method attains 92.6% attribute editing accuracy, improves fidelity by 37%, and processes images at 50 ms per image. Its core contributions are: (i) eliminating reliance on manual annotations or model fine-tuning; and (ii) establishing the first interpretable, composable, and lightweight unsupervised editing paradigm for StyleGAN.

Technology Category

• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Problem

• Image Style Transfer
• Computational Efficiency
• Feature Adaptation

Innovation

• StyleAutoEncoder
• efficient feature tuning
• flexible neural network design