DIFFUSION AUTOENCODERS ARE SCALABLE IMAGE TOKENIZERS

Summary from the TLDR AI newsletter

The modern workhorse of multimodal understanding and generation is learned Tokenizers. These models are typically autoencoder style with a learned discrete codebook. They tend to work well but are very hard to train and require careful tuning of many auxiliary losses. This work shows that with a single diffusion loss, image tokenization is stable, scalable, and higher quality than many traditional methods.