AttnGAN: Fine-Grained Text to Image Generation with Attentional Generative Adversarial Networks

Overview

Explore the innovative AttnGAN model for fine-grained text-to-image generation in this 46-minute lecture from the University of Central Florida. Delve into the architecture's key components, including the text encoder, conditioning augmentation, generator, attention network, and image encoder. Examine the DAMSM loss and its role in improving image quality. Learn about experimental results on various datasets, evaluation metrics like Inception score, and component analysis. Discover the model's capabilities in generating novel scenarios and understand its limitations in capturing global coherent structure. Gain insights into the challenges and advancements in text-to-image synthesis using attentional generative adversarial networks.

Syllabus

Intro
Problem: Text-to-image
Related work
Architecture - Motivation
Architecture - Text Encoder
Architecture - Conditioning Augmentation
Architecture - Generator F.
Architecture - Attention network Fatin
Architecture - Image Encoder
Architecture - DAMSM loss
Experiments - Datasets
Experiments - Evaluation • Inception score
Experiments - Component Analysis
Experiments - Qualitative (CUB)
Experiments - Novel scenarios
Experiments - Failure cases Did not capture global coherent structure