Cold Diffusion from Scratch (github.com)

A new project called "Cold Diffusion: Inverting Arbitrary Image Transforms Without Noise" has been unveiled at NeurIPS 2023, showcasing a novel approach to image processing. Developed from scratch, the initiative demonstrates how cold diffusion can effectively invert complex image transforms without introducing noise, a frequent issue in traditional image manipulation techniques. The project is accessible via GitHub, where users can clone the repository and quickly set up the environment to train the model using a sample dataset, FashionMNIST. This development is significant for the AI/ML community as it pushes the boundaries of how we understand and implement image transformation and recovery methods. By providing qualitative results from training on FashionMNIST for approximately 20 epochs, the project reveals the efficacy of using a noise-mitigation strategy by utilizing an initial seed-noise distribution derived from randomly blurred validation samples. This could influence various applications such as image restoration, computer vision, and generative models, positioning cold diffusion as a potentially robust technique for future research and development in the field.