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

TL;DR

AttnGAN introduces an attention-driven, multi-stage generative adversarial network that produces detailed images from text descriptions by focusing on relevant words at different image regions, significantly improving image quality.

Contribution

The paper presents a novel attentional generative network and a fine-grained image-text matching loss, advancing text-to-image synthesis with detailed, region-specific image generation.

Findings

Achieved 14.14% higher inception score on CUB dataset

Boosted performance by 170.25% on COCO dataset

Visualized attention layers to show word-level image region focus

Abstract

In this paper, we propose an Attentional Generative Adversarial Network (AttnGAN) that allows attention-driven, multi-stage refinement for fine-grained text-to-image generation. With a novel attentional generative network, the AttnGAN can synthesize fine-grained details at different subregions of the image by paying attentions to the relevant words in the natural language description. In addition, a deep attentional multimodal similarity model is proposed to compute a fine-grained image-text matching loss for training the generator. The proposed AttnGAN significantly outperforms the previous state of the art, boosting the best reported inception score by 14.14% on the CUB dataset and 170.25% on the more challenging COCO dataset. A detailed analysis is also performed by visualizing the attention layers of the AttnGAN. It for the first time shows that the layered attentional GAN is able to automatically select the condition at the word level for generating different parts of the image.