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

**Authors:** Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, Xiaodong He  
**arXiv:** [1711.10485](https://arxiv.org/abs/1711.10485)  
**Date:** 2017-11-28  
**Summary page:** https://www.summarizepaper.com/en/paper/1711.10485v1/

## Key points

- AttnGAN is a new type of Generative Adversarial Network (GAN) that generates images from text descriptions.
- It uses attention-driven, multi-stage refinement to synthesize fine-grained details in different subregions of the image.
- The network pays attention to relevant words in the natural language description and uses a deep attentional multimodal similarity model to compute an image-text matching loss for training the generator.
- AttnGAN outperforms previous state-of-the-art models, with a 14.14% boost on the CUB dataset and 170.25% on the more challenging COCO dataset.
- Visualizing the attention layers of AttnGAN shows that it can automatically select conditions at the word level for generating different parts of an image, setting it apart from other generative models.

## Summary

AttnGAN is a novel Attentional Generative Adversarial Network that enables attention-driven, multi-stage refinement for fine-grained text-to-image generation. The network synthesizes fine-grained details at different subregions of the image by paying attention to the relevant words in the natural language description. Additionally, a deep attentional multimodal similarity model computes a fine-grained image-text matching loss for training the generator. AttnGAN outperforms previous state-of-the-art models, 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 performed by visualizing the attention layers of AttnGAN, which shows that it can automatically select conditions at the word level for generating different parts of an image. This capability has not been demonstrated before and sets AttnGAN apart from other generative models in its ability to generate high quality images with fine grained details based on natural language descriptions.