GenFirst

About

GenFirst is a ByteDance Seed release for training a latent VAE jointly with a generative prior. A two-phase weighting schedule lets the prior shape the latent space first, then lets the decoder recover detail, which avoids the latent collapse that makes end-to-end training unstable. The release includes end-to-end VAEs and SiT and MMDiT priors for class-conditional and text-to-image generation, and the weights are limited to non-commercial research use.