Multi-Label Conditional Generation From Pre-Trained Models.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Modern generative models excel in data quality but struggle with attribute control.
• Generating data with specific properties (e.g., age, molecular weight) remains a challenge.

Purpose of the Study:

• To introduce PluGeN (Plugin Generative Network), a novel technique to enhance existing generative models.
• To enable precise attribute control and manipulation in generated data.
• To facilitate the generation of samples with rare or unseen attribute combinations.

Main Methods:

• PluGeN transforms entangled latent representations into a multi-dimensional space using a flow-based module.
• Each attribute is modeled as an independent one-dimensional distribution, enabling disentanglement.
• The approach was integrated with Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs).

Main Results:

• PluGeN successfully enabled conditional generation and manipulation of images, chemical molecules, and 3D point clouds.
• The method allows for the creation of samples with desired attributes and manipulation of existing ones.
• Generated samples included novel attribute combinations like 'young person with gray hair' or 'men with make-up'.

Conclusions:

• PluGeN offers a flexible and effective plugin for pre-trained generative models, overcoming attribute control limitations.
• The disentangled representation facilitates attribute manipulation and the generation of diverse, even unseen, data variations.
• The technique's ability to train on partially labeled data enhances its practical applicability across various domains.