1 paper
Jiawei Zhang, Shuhao Liu, Rong Huang +4
One-dimensional (1D) variable-length visual tokenizers enable adaptive compression by varying the number of tokens, allowing downstream autoregressive (AR) models to flexibly trade…