AI and Machine Learning

Generative Models

Learning to sample from a distribution rather than to label one.

Colour is the family; a dashed line is the second member of it.

Drag to pan · scroll to zoom · click a node to open it
G classifier-free-guidance Classifier-Free Guidance diffusion-model Diffusion Model classifier-free-guidance->diffusion-model conditioning alone is too weak to follow a prompt generative-adversarial-network Generative Adversarial Network diffusion-model->generative-adversarial-network a stable objective instead of a minimax game that collapses neural-denoising Learned Denoising diffusion-model->neural-denoising denoises repeatedly, from pure noise, to synthesise variational-autoencoder Variational Autoencoder generative-adversarial-network->variational-autoencoder an adversarial critic instead of a likelihood bound latent-diffusion Latent Diffusion latent-diffusion->diffusion-model denoising at pixel resolution is prohibitively expensive latent-diffusion->variational-autoencoder the latent space it works in comes from one monte-carlo-integration Monte Carlo Integration variational-autoencoder->monte-carlo-integration the bound is estimated by sampling the latent

5 nodes

Variational Autoencoder

Learns an encoder to a latent distribution and a decoder back, trained on a lower bound of the likelihood. Gives a latent space you can interpolate i… · 2013

Generative Adversarial Network

Trains a generator against a discriminator that tries to tell real from fake. Produces sharp samples, and the minimax objective is notoriously unstab… · 2014

Diffusion Model

Learns to reverse a gradual noising process, generating by denoising repeatedly from pure noise. Trades sampling speed for stable training and far be… · 2020

Latent Diffusion

Runs the diffusion process in the compressed latent space of an autoencoder rather than on pixels. The reason text-to-image generation fits on consum… · 2021

Classifier-Free Guidance

Trains one model both conditioned and unconditioned, then extrapolates away from the unconditional prediction at sampling time. Buys prompt adherence… · 2022