TC-AE: Unlocking Token Capacity for Deep Compression Autoencoders

Summary

TC-AE is a ViT-based deep compression autoencoder architecture introduced in an arXiv paper (2504.06852) by Teng Li, Ziyuan Huang, and Cong Chen. Existing autoencoder methods typically increase the channel count of latent representations to preserve reconstruction quality at high compression ratios, but this approach often causes latent representation collapse. TC-AE tackles this challenge from the token-space perspective: it decomposes token-to-latent compression into two stages and enhances the semantic structure of image tokens through joint self-supervised training. As a result, the model achieves substantially improved reconstruction and generation performance under deep compression. This work is relevant to computer vision research, particularly for latent-space generative models that rely on compact, high-fidelity latent representations.