Diffusion model becomes a very popular model not just in image generation, but also in protein design.
Let's start with a straightforward scenario to study the diffusion model.
The Denoising Diffusion Probabilistic Model (DDPM) comprises two integral processes: the forward process and the reverse process.
Forward Process: In this process, noise sampled from a Gaussian distribution is incrementally added to an image at each time step. This forward process iteratively refine the pixel distribution of the image until it approximates a Gaussian distribution.
Reverse Process: Conversely, the reverse process involves predicting the noise and subsequently removing it from a noisy image at each time step.
