ByteDance Seed releases GenFirst: train the VAE and generator together, generation first, reconstruction second
Official source · auto-listedOriginal title ByteDance-Seed/GenFirst
Image generators usually train a VAE first to compress images into a latent space, then a generator on top. ByteDance Seed's GenFirst trains both together: generation shapes the latent space first, and reconstruction is phased in to avoid latent collapse. On ImageNet-256, SiT reaches gFID 0.97 with CFG. Only the EiT version is out; weights and some code are CC BY-NC 4.0, non-commercial only.
Why it matters
The paper came out in August; this release opens code and weights for reproduction. VAEs and generators are usually trained in two stages; GenFirst trains them jointly end to end. Weights and some code are non-commercial; only EiT is out.
Sources3
Auto-listed by dalao.ai: the headline is a fixed template and the original title is shown as written. The official page is the record; its publisher keeps authorship.