✳flâneur — a map of the web's best reading

Bare-bones Diffusion Models

madebyoll.in · 2,384 words · saved by 2 readers

To understand how diffusion models generate images, I wrote code for training a bare-bones diffusion model without all the fanciness (no “epsilon parameterization”, no “Gaussian conditionals”, no sqrt_one_minus_alphas_cumprod). Based on that code, I trained a tiny diffusion model that generates 512×512 botanical images in your web browser. You can play with that demo here. So... how do diffusion models generate images? Why are these images as good as they are, and not better (or worse)? And... if we’re talking about conjuring new images from nothingness, what does “better” even mean? The image generation task is: given only some random noise as input, generate new images. Furthermore, generated images should look “real” (when compared to some reference dataset of, you know, actually-real images). Ideally, we want an image generator to exhibit: An archetypal “boring” task is image classification. It feels contrived (just look at those categories) and it also feels too easy (even a child

Bare-bones Diffusion Models Bare-bones Diffusion Models 05 February 2023 To understand how diffusion models generate images, I wrote code for training a bare-bones diffusion model without all the fanciness (no "epsilon parameterization", no "Gaussian conditionals", no sqrt_one_minus_alphas_cumprod ). Based on that code, I trained a tiny diffusion model that generates 512×512 botanical images in your web browser. You can play with that demo here . So... how do diffusion models generate images? Why are these images as good as they are, and not better (or worse)? And... if we're talking abo

Explore this link on the map →

saved by

related reading