Scaling the Codebook Size of VQGAN to 100,000 with a Utilization Rate of 99%

Benchmark Model Rank Results
image-reconstruction-on-imagenetVQGAN-LC (16x16)#21FID: 2.62LPIPS: 0.120PSNR: 23.80SSIM: 0.589
image-generation-on-imagenet-256x256VQGAN-LC + LDM (32x32 tokens)#185FID: 4.81