Tag
IDEAL proposes an in-depth alignment framework for discrete representation autoencoding, jointly aligning quantized tokens with shallow and deep VFM features to achieve superior reconstruction and generation performance.