# Synthetic Data from Diffusion Models Improves ImageNet Classification

**Authors:** Shekoofeh Azizi, Simon Kornblith, Chitwan Saharia, Mohammad Norouzi, David J. Fleet  
**arXiv:** [2304.08466](https://arxiv.org/abs/2304.08466)  
**Date:** 2023-04-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.08466v1/

## Key points

- Deep generative models have advanced in generating high-quality, photo-realistic images based on text prompts.
- These models can be used for generative data augmentation to enhance challenging discriminative tasks.
- Large-scale text-to-image diffusion models can be fine-tuned to produce class conditional models with state-of-the-art Frechet Inception Distance (FID) and Inception Score at a resolution of 256x256.
- Generated samples achieve a new state-of-the-art in Classification Accuracy Scores, with 64.96 for 256x256 generative samples and improving to 69.24 for 1024x1024 samples.
- Augmenting the ImageNet training set with these generated samples leads to significant improvements in ImageNet classification accuracy compared to strong ResNet and Vision Transformer baselines.
- Previous studies have shown that synthetic data generated with GLIDE improves zero-shot and few-shot image classification performance.
- Fine-tuning the Imagen text-to-image model for class conditional ImageNet leads to state-of-the-art models.
- This research highlights the potential of using large-scale text-to-image diffusion models for generative data augmentation, leading to improved performance in challenging discriminative tasks such as ImageNet classification.

## Summary

Deep generative models have made significant advancements in generating high-quality, photo-realistic images based on text prompts. These models have the potential to be used for generative data augmentation, which can enhance challenging discriminative tasks. In this study, the researchers demonstrate that large-scale text-to-image diffusion models can be fine-tuned to produce class conditional models with state-of-the-art Frechet Inception Distance (FID) and Inception Score at a resolution of 256x256. The results show that the generated samples achieve a new state-of-the-art in Classification Accuracy Scores, with 64.96 for 256x256 generative samples and improving to 69.24 for 1024x1024 samples. By augmenting the ImageNet training set with these generated samples, significant improvements in ImageNet classification accuracy are observed compared to strong ResNet and Vision Transformer baselines. The authors provide additional context by discussing related work in synthetic data generation using diffusion models. Previous studies have shown that synthetic data generated with GLIDE improves zero-shot and few-shot image classification performance. Augmenting individual images using a pretrained diffusion model has also demonstrated improvements in few-shot settings. Two recent papers have trained ImageNet classifiers on images generated by diffusion models but did not fine tune them. However, these studies found that the generated images did not improve accuracy on the clean ImageNet validation set. In contrast, this study shows that fine tuning the Imagen text to image model for class conditional ImageNet leads to state of the art models. Overall, this research highlights the potential of using large scale text to image diffusion models for generative data augmentation, leading to improved performance in challenging discriminative tasks such as ImageNet classification.