gFID

FuseReg: Layer-Fusion Regularization for Bridging the Reconstruction-Generation Gap in Representation Autoencoders

FuseReg treats layer fusion as a trainable distribution, making latent codes robust to varied encoder subsets …