ML//diffusion model

A diffusion model is a generative model that learns to reverse a noising process: data is corrupted step by step until only static remains, a neural net is trained to undo each step, and generation is then iterative denoising from random noise. Training is stable (no mode collapse) and the outputs are of high quality, but inference is slow because it takes many steps. On those terms it replaced the GAN for image generation: more stable training and better quality, at a cost in speed of orders of magnitude.

A diffusion model is a generative model that learns to reverse a noising process: data is corrupted step by step until only static remains, a neural net is trained to undo each step, and generation is then iterative denoising from random noise. Training is stable (no mode collapse) and the outputs are of high quality, but inference is slow because it takes many steps. On those terms it replaced the GAN for image generation: more stable training and better quality, at a cost in speed of orders of magnitude.