TC-AE: Unlocking Token Capacity for Deep Compression Autoencoders

Summary

TC-AE is a ViT-based deep compression autoencoder proposed by Teng Li, Ziyuan Huang, and Cong Chen in an arXiv paper (2504.06852) released on April 9, 2025. The paper addresses a key limitation in existing deep compression autoencoders: these methods typically increase the channel count of latent representations to maintain reconstruction quality at high compression ratios, which often causes latent representation collapse. Instead of expanding channels, TC-AE tackles the problem from a token-space perspective. It decomposes the token-to-latent compression process into two stages and enhances the semantic structure of image tokens through joint self-supervised training. According to the paper, this design enables significantly improved reconstruction and generation performance under deep compression. The work is relevant to image generation and latent-space modeling, where compact yet informative latent representations are critical for the efficiency and quality of downstream generative models. Full details are available on arXiv.