# CogView: Mastering Text-to-Image Generation via Transformers

**Authors:** Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, Jie Tang  
**arXiv:** [2105.13290](https://arxiv.org/abs/2105.13290)  
**Date:** 2021-05-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2105.13290v3/

## Key points

- A team of researchers led by Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang introduce a groundbreaking approach called CogView.
- CogView utilizes a 4-billion-parameter Transformer model to advance text-to-image generation capabilities.
- The approach showcases versatility through various finetuning strategies for downstream tasks such as style learning, super-resolution techniques, text-image ranking mechanisms, and applications in fashion design.
- Methods are introduced to enhance pretraining stability by addressing issues like eliminating NaN losses.
- CogView achieves state-of-the-art performance on the blurred MS COCO dataset in terms of Fréchet Inception Distance (FID), surpassing previous models based on Generative Adversarial Networks (GANs) and outperforming DALL-E.

## Summary

In the realm of  a longstanding challenge has been the development of a robust generative model coupled with a deep understanding of  Addressing this issue head-on, a team of researchers led by Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang introduce  This groundbreaking approach harnesses the power of a 4-billion-parameter Transformer equipped with a  to push the boundaries of text-to-image generation. CogView not only presents an innovative solution to this complex problem but also showcases its versatility through various finetuning strategies for downstream tasks. These include style learning, super-resolution techniques, text-image ranking mechanisms, and even applications in fashion design. Moreover, the team introduces methods to enhance pretraining stability by addressing issues such as eliminating NaN losses. One notable achievement of CogView is its state-of-the-art performance on the blurred MS COCO dataset in terms of Fréchet Inception Distance (FID). By surpassing previous models based on Generative Adversarial Networks (GANs) and even outperforming DALL-E—a recent work with similar objectives—CogView establishes itself as a frontrunner in the field of text-to-image generation.