GenFirst: Generation Before Reconstruction for Stable End-to-End Latent Generative Modeling

Benchmark Model Rank Results
image-generation-on-imagenet-256x256GenFirst EiT (675M, 800 epochs)–FID: 0.969
image-generation-on-imagenet-512x512GenFirst EiT (675M, 400 epochs)–FID: 1.23Inception score: 286.1
text-to-image-generation-on-dpg-benchGenFirst EiT (1.3B, Qwen3-1.7B text encoder)–DPG-Bench Score: 84.65
text-to-image-generation-on-genevalGenFirst EiT1.3B CLIPtextencoder–Overall: 0.9