Lecture 002

Autoencoder: just autoencoder

Denoising Autoencoder: autoencoder, but predict noise from corrupted

Variational Autoencoder: just add variation

VQ-VAE (Vector Quantization): maps continuous vectors to a codebook of discrete set of vectors

VQ-GAN: VQ-VAE with Transformer and GAN

Evaluating Data Attribution for Text-to-Image Models

Problem: what are the training images and corresponding scores that contribute in the generation process of this image

Solution: get a bunch of dataset by fine-tuning 1000 models with 1000 images. And then train a CLIP-like model.

Understanding and Mitigating Copying in Diffusion Models