Recent works on learned image compression perform encoding and decoding
processes in a full-resolution manner, resulting in two problems when deployed
for practical applications. First, parallel acceleration of the autoregressive
entropy model cannot be achieved due to serial decoding.