# Cold Diffusion: Inverting Arbitrary Image Transforms Without Noise

**Authors:** Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie S. Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, Tom Goldstein  
**arXiv:** [2208.09392](https://arxiv.org/abs/2208.09392)  
**Date:** 2022-08-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2208.09392v1/

## Key points

- Diffusion models in image processing traditionally involve adding Gaussian noise and using an image restoration operator.
- The authors discovered that the generative behavior of diffusion models is not heavily influenced by the type of image degradation used.
- A diverse range of generative models can be created by varying the type of degradation employed.
- Deterministic degradations like blur or masking can be used instead of stochastic noise for high-quality image generation.
- Noise is not essential for diffusion models to function effectively in gradient Langevin dynamics or variational inference.
- Fully deterministic models capable of inverting arbitrary processes without relying on noise-based techniques have been introduced, opening up new possibilities for generalized diffusion models.
- The availability of code on GitHub facilitates exploration and implementation of these innovative approaches.

## Summary

The paper "Cold Diffusion: Inverting Arbitrary Image Transforms Without Noise" by Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie S. Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein delves into the concept of diffusion models in image processing. These models traditionally involve adding Gaussian noise to images and using an image restoration operator to reverse the degradation. However, the authors make a groundbreaking discovery that the generative behavior of diffusion models is not heavily influenced by the specific type of image degradation used. They demonstrate that a diverse range of generative models can be created simply by varying the type of degradation employed. Furthermore, they show that even deterministic degradations like blur or masking can be used instead of stochastic noise in training and testing procedures for high-quality image generation. This challenges the conventional understanding within the research community that noise is essential for diffusion models to function effectively in gradient Langevin dynamics or variational inference. By introducing fully deterministic models capable of inverting arbitrary processes without relying on noise-based techniques, this research opens up new possibilities for generalized diffusion models in image processing. The availability of their code on GitHub also facilitates exploration and implementation of these innovative approaches. Through their work on "Cold Diffusion," Bansal et al. have significantly advanced our understanding of generative modeling and paved the way for more versatile and efficient image restoration techniques.