# Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding

**Authors:** Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, Mohammad Norouzi  
**arXiv:** [2205.11487](https://arxiv.org/abs/2205.11487)  
**Date:** 2022-05-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2205.11487v1/

## Key points

- Authors present Imagen, a text-to-image diffusion model achieving photorealism and language understanding
- Imagen combines large transformer language models with diffusion models for image generation
- Generic large language models like T5 effectively encode text for image synthesis in Imagen
- Increasing size of language model has greater impact on fidelity and alignment than increasing image diffusion model size
- Imagen achieves state-of-the-art FID score of 7.27 on COCO dataset without being trained on COCO data itself
- DrawBench introduced as comprehensive benchmark for evaluating text-to-image models
- Imagen emerges as preferred choice over recent methods in terms of sample quality and image-text alignment using DrawBench evaluation metrics and human raters' opinions
- Imagen demonstrates advancements in photorealistic text to image synthesis by leveraging large language models and diffusion models

## Summary

In their paper titled "Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding," authors Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho and David J Fleet and Mohammad Norouzi present Imagen—a text-to-image diffusion model that achieves an unprecedented level of photorealism and language understanding. The model combines the power of large transformer language models in text comprehension with diffusion models for high-fidelity image generation. The authors make a significant discovery that generic large language models like T5 can effectively encode text for image synthesis in Imagen. They find that increasing the size of the language model has a greater impact on sample fidelity and image-text alignment than increasing the size of the image diffusion model. As a result of this finding Imagen achieves a state-of-the-art FID score of 7.27 on the COCO dataset without being trained on COCO data itself. To evaluate text-to-image models comprehensively the authors introduce DrawBench—a comprehensive and challenging benchmark specifically designed for this purpose. In comparison to recent methods such as VQ-GAN+CLIP Latent Diffusion Models and DALL-E 2 using DrawBench evaluation metrics and human raters' opinions in side by side comparisons—Imagen emerges as the preferred choice due to its superior sample quality and image text alignment. Overall Imagen demonstrates remarkable advancements in photorealistic text to image synthesis by leveraging large language models and diffusion models. The paper provides valuable insights into improving sample fidelity and aligning images with textual descriptions. For more detailed information about their results and findings from DrawBench evaluations readers are directed to the official website of Imagen at https://imagen.research.google/.