# Zero-Shot Text-to-Image Generation

**Authors:** Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, Ilya Sutskever  
**arXiv:** [2102.12092](https://arxiv.org/abs/2102.12092)  
**Date:** 2021-02-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2102.12092v1/

## Key points

- Traditional text-to-image generation focuses on improving modeling assumptions for fixed datasets
- Recent research suggests scaling up model size and data can lead to better results and improved generalization
- Researchers propose a simple approach based on an autoregressive transformer that models text and image tokens as a single stream of data
- Model is capable of generalizing in unexpected ways, composing unusual concepts at high levels of abstraction, and performing variable binding
- Demonstrates zero-shot image-to-image translation control by natural language, including transformations like changing colors or styles
- When executed at scale with 12 billion parameters trained on 250 million images, the model outperforms previous domain-specific approaches in terms of zero-shot performance and range of capabilities
- Findings suggest scaling up may be a useful driver for progress in text-to-image generation tasks.

## Summary

The field of text-to-image generation has traditionally focused on improving modeling assumptions for training on a fixed dataset. Recent research suggests that scaling up the model size and data can lead to improved generalization and better results. In this study, the researchers propose a simple approach based on an autoregressive transformer that models text and image tokens as a single stream of data. The model is capable of generalizing in unexpected ways, such as composing unusual concepts at high levels of abstraction and performing variable binding. It also demonstrates zero-shot image-to-image translation control by natural language, including transformations like changing colors or styles. When executed at scale with 12 billion parameters trained on 250 million images, the model outperforms previous domain-specific approaches in terms of zero-shot performance and range of capabilities. The findings suggest that scaling up may be a useful driver for progress in text-to-image generation tasks.