Exploring multi-semantic disentangled controls in GANs using conjugate gradient optimization

Zhenbo Yu

doi:10.1016/j.patrec.2025.01.018

Return

Exploring multi-semantic disentangled controls in GANs using conjugate gradient optimization

2025-03-01

PRE

DOI：10.1016/j.patrec.2025.01.018

Original

Original request for help

Save

Abstract

En 中文

In the latent space of Generative Adversarial Networks (GANs), numerous semantic vectors exist. Discovering meaningful semantic vectors associated with certain semantics in the generated image is crucial, as they can be manipulated to achieve effective image editing. Previous approaches typically utilize manual annotation to train classifiers for identifying different semantics. However, these methods often prioritize the feasibility of image editing but fail to achieve the disentanglement of different attributes in edited images. To address this issue, we propose a quadratic conjugate gradient optimization method called QCGO. This method effectively mitigates the entanglement between different semantic attributes by leveraging the framework of quadratic optimization problems. QCGO not only significantly enhances the effect of disentanglement between different semantics in images but also accelerates convergence speed to rapidly obtain meaningful semantics. Concretely, we first formulate image editing tasks as a quadratic optimization problem. Next, we employ the conjugate gradient method to identify meaningful semantics, leveraging the conjugate property between different conjugate vectors to better disentangle various attributes in images. Extensive experiments demonstrate that our QCGO method achieves satisfactory effects of disentanglement in image editing and exhibits faster convergence speed compared to other state-of-the-art methods.

Keywords:

Generative Adversarial Network

Conjugate gradient optimization

Semantic editing