Quality (philosophy)

Quantum mechanics

Philosophy

Computer security

Generative adversarial network

Astrophysics

Power (physics)

Computer science

Visual arts

Acoustics

Interpolation (computer graphics)

Variation (astronomy)

Architecture

Physics

Identity (music)

Generative grammar

Image (mathematics)

Epistemology

Artificial intelligence

Generator (circuit theory)

Deep learning

Machine learning

Adversarial system

Key (lock)

IEEE Transactions on Pattern Analysis and Machine Intelligence

We propose an alternative generator architecture for generative adversarial networks, borrowing from style transfer literature. The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis. The new generator improves the state-of-the-art in terms of traditional distribution quality metrics, leads to demonstrably better interpolation properties, and also better disentangles the latent factors of variation. To quantify interpolation quality and disentanglement, we propose two new, automated methods that are applicable to any generator architecture. Finally, we introduce a new, highly varied and high-quality dataset of human faces.

A Style-Based Generator Architecture for Generative Adversarial Networks