GenFirst: Generation Before Reconstruction for Stable End-to-End Latent Generative Modeling
The paper claims stable one-pass training for latent image generators by letting generation shape the latent space before reconstruction tightens it.
GenFirst targets the collapse problem that appears when autoencoding and generation are optimized together. The authors argue that KL entropy keeps the latent posterior from degenerating, while early reconstruction pressure can overpower the slower generative objective. Their schedule starts with weak reconstruction, then strengthens it to restore detail after the latent space is more generation-friendly. Reported results include SiT gFID of 0.97 with CFG on ImageNet-256 and MMDiT GenEval of 0.90 for text-to-image generation. Source: HF Daily Papers' note.
GenFirst targets the collapse problem that appears when autoencoding and generation are optimized together. The authors argue that KL entropy keeps the latent posterior from degenerating, while early reconstruction pressure can overpower the slower generative objective. Their schedule starts with weak reconstruction, then strengthens it to restore detail after the latent space is more generation-friendly. Reported results include SiT gFID of 0.97 with CFG on ImageNet-256 and MMDiT GenEval of 0.90 for text-to-image generation. Source: HF Daily Papers' note.
score 5