A text-to-image GAN design that injects frozen pretrained vision models into adversarial training to rival diffusion-model image quality while retaining a GAN's fast single-forward-pass generation. Its two novel parts are a vision-empowered discriminator that stacks multiple frozen pretrained encoders (e.g. CLIP, DINO) from a Pretrained Vision Model Bank to supply richer perceptual gradients, and a high-capacity generator of stacked High-capacity Fusion Blocks (deep text-image fusion modules plus a global fusion module) for tighter text-image alignment. Reaches FID 5.48 on COCO and 9.38 on CUB at roughly 120x faster inference than diffusion baselines.
TIGER (Text-to-Image GAN with pretrained rEpresentations) is a text-to-image generative adversarial network from Xiaozhou You and Jian Zhang (Peking University) that injects frozen pretrained vision …