# GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models

**Authors:** Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, Mark Chen  
**arXiv:** [2112.10741](https://arxiv.org/abs/2112.10741)  
**Date:** 2021-12-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2112.10741v1/

## Key points

- Diffusion models are used for text-conditional image synthesis
- Two guidance strategies compared: CLIP guidance and classifier-free guidance
- Classifier-free guidance preferred by evaluators in terms of photorealism and caption similarity
- Even with CLIP reranking, samples from a 3.5 billion parameter text-conditional diffusion model using classifier-free guidance are favored by evaluators
- Models can be fine-tuned for image inpainting, enabling text-driven image editing capabilities
- Code and weights provided for the implementation of a smaller model trained on a filtered dataset
- Diffusion models show potential in generating high-quality synthetic images guided by text inputs
- Classifier-free guidance is more effective than CLIP guidance in terms of photorealism and caption similarity
- Fine-tuning allows for image inpainting and expands text-driven image editing possibilities

## Summary

The paper titled "GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models" explores the use of diffusion models for text-conditional image synthesis. Diffusion models have been proven to generate high-quality synthetic images, particularly when combined with a guidance technique that balances diversity and fidelity. The authors compare two different guidance strategies: CLIP guidance and classifier-free guidance. Through human evaluation, it is found that the classifier-free guidance strategy is preferred by evaluators in terms of both photorealism and caption similarity. This strategy also tends to produce photorealistic samples. Interestingly, even when the expensive CLIP reranking method is employed by DALL-E, samples from a 3.5 billion parameter text-conditional diffusion model using classifier-free guidance are favored by human evaluators. Furthermore, the authors discover that their models can be fine-tuned for image inpainting, enabling powerful text-driven image editing capabilities. They train a smaller model on a filtered dataset and provide the code and weights for this implementation. Overall, this research demonstrates the potential of diffusion models in generating high-quality synthetic images guided by text inputs. The findings highlight the effectiveness of classifier-free guidance over CLIP guidance in terms of photorealism and caption similarity. Additionally, the study showcases the ability to perform image inpainting through fine-tuning, expanding the range of text-driven image editing possibilities.