A Live Benchmark of Diffusion VAE!

This page is a live benchmark of diffusion VAE benchmark in our paper Making Reconstruction FID Predictive of Diffusion Generation FID.

[arxiv][github code] [huggingface pre-trained models]

We compare gFID of SiT trained on ImageNet 256 for 40 epochs, vs PSNR and our proposed iFID metric.

Now 15 VAEs are included! If you want your VAE to be included, submit an issue in Github.