Ziyuan Huang

Doubling the number of tokens in a ViT-based autoencoder, combined with staged compression and self-supervised pretraining, dramatically improves generative performance under deep compression, without increasing the latent budget.