GenFirst addresses latent collapse in end-to-end training of latent generative models by reversing the usual two-stage pipeline: a generation-before-reconstruction strategy lets the generative objective shape the latent space under weak reconstruction pressure before reconstruction is progressively strengthened. Tested with multiple generative priors, the approach reaches a gFID of 0.97 with classifier-free guidance and 1.45 without it on ImageNet-256.
