# Adversarial Diffusion Distillation

**Authors:** Axel Sauer, Dominik Lorenz, Andreas Blattmann, Robin Rombach  
**arXiv:** [2311.17042](https://arxiv.org/abs/2311.17042)  
**Date:** 2023-11-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.17042v1/

## Key points

- The authors propose a training approach called Adversarial Diffusion Distillation (ADD) for efficient sampling of large-scale foundational image diffusion models in 1-4 steps while maintaining high image quality.
- ADD enables single-step, real-time image synthesis with foundation models.
- Performance evaluation of ADD was done through user preference studies, assessing prompt adherence and overall image quality.
- ADD-XL outperforms LCM-XL (4 steps) with just a single step and can beat SDXL (50 steps) with only four steps in most comparisons, making it the state-of-the-art method in both single and multiple step settings.
- Qualitative results showcase the capabilities of ADD-XL, including enhanced realism and reduced oversmoothing compared to diffusion model samples.
- In terms of speed, Fig. 7 visualizes the inference speeds relative to ELO scores, providing insights into performance considering both speed and image quality.
- Table 2 compares different few-step sampling and distillation methods using the same base model; ADD outperforms all other approaches, including the standard DPM solver with eight steps.
- The proposed ADD approach is highly effective for sampling large-scale foundational image diffusion models in a few steps while maintaining high image quality; it surpasses existing methods and reaches the level of state-of-the-art diffusion models with only four steps.
- Code and weights are available on GitHub and Hugging Face for further exploration and implementation.

## Summary

In their paper, titled "Adversarial Diffusion Distillation," the authors propose a novel training approach called Adversarial Diffusion Distillation (ADD) that allows for efficient sampling of large-scale foundational image diffusion models in just 1-4 steps while maintaining high image quality. The goal of their approach is to enable single-step, real-time image synthesis with foundation models. To evaluate the performance of ADD, the authors conducted a quantitative comparison with other approaches using user preference studies rather than automated metrics. They assessed both prompt adherence and overall image quality and computed win percentages for pairwise comparisons and ELO scores when comparing several approaches. The reported ELO scores represent the mean scores between prompt following and image quality. The study results showed that ADD-XL outperforms LCM-XL (4 steps) with just a single step. Additionally, ADD-XL can beat SDXL (50 steps) with only four steps in the majority of comparisons, making it the state-of-the-art method in both single and multiple step settings. To complement their quantitative studies, the authors also provided qualitative results. They presented additional samples and qualitative comparisons to showcase the capabilities of ADD-XL. The adversarial loss used in ADD-XL enhances realism by improving textures such as fur, fabric, and skin while reducing oversmoothing commonly observed in diffusion model samples. However, it was noted that ADD-XL's overall sample diversity tends to be lower compared to its teacher model SDXL. In terms of speed, Fig. 7 visualizes the inference speeds of different models relative to their ELO scores which provides insights into how well each model performs considering both speed and image quality. Furthermore, Table 2 compares different few-step sampling and distillation methods using the same base model; results demonstrate that ADD outperforms all other approaches including the standard DPM solver with eight steps. Overall, the proposed Adversarial Diffusion Distillation (ADD) approach proves to be a highly effective method for sampling large scale foundational image diffusion models in just a few steps while maintaining high image quality; its performance surpasses existing methods and reaches the level of state–of–the–art diffusion models with only four steps. The authors have made code and weights available on GitHub and Hugging Face for further exploration and implementation.