# Diffusion Models already have a Semantic Latent Space

**Authors:** Mingi Kwon, Jaeseok Jeong, Youngjung Uh  
**arXiv:** [2210.10960](https://arxiv.org/abs/2210.10960)  
**Date:** 2022-10-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2210.10960v2/

## Key points

- Diffusion models have achieved outstanding generative performance across various domains
- However, they lack a refined semantic latent space crucial for controlling the generative process
- Mingi Kwon, Jaeseok Jeong, and Youngjung Uh propose an asymmetric reverse process (Asyrp) to address this issue
- Asyrp discovers the semantic latent space in frozen pretrained diffusion models
- Their proposed semantic latent space, named h-space, has several desirable properties for accommodating semantic image manipulation such as homogeneity, linearity, robustness and consistency across timesteps
- The authors introduce a principled design of the generative process that allows for versatile editing and quality boosting by quantifiable measures such as editing strength of an interval and quality deficiency at a timestep
- The Asyrp method is applicable to various architectures such as DDPM++, iD-DPM and ADM as well as datasets including CelebA-HQ, AFHQ-dog, LSUN-church, LSUN-bedroom and METFACES.
- The authors' project page provides further details on their work.

## Summary

Diffusion models have been successful in achieving outstanding generative performance across various domains. However, they lack a refined semantic latent space that is crucial for controlling the generative process. To address this issue, Mingi Kwon, Jaeseok Jeong, and Youngjung Uh propose an asymmetric reverse process (Asyrp) that discovers the semantic latent space in frozen pretrained diffusion models. Their proposed semantic latent space, named h-space, has several desirable properties for accommodating semantic image manipulation such as homogeneity, linearity, robustness and consistency across timesteps. The authors also introduce a principled design of the generative process that allows for versatile editing and quality boosting by quantifiable measures such as editing strength of an interval and quality deficiency at a timestep. The Asyrp method is applicable to various architectures such as DDPM++, iD-DPM and ADM as well as datasets including CelebA-HQ, AFHQ-dog, LSUN-church, LSUN-bedroom and METFACES. The authors' project page provides further details on their work.