Soft Truncation: A Universal Training Technique of Score-based Diffusion Model for High Precision Score Estimation

Open paper
Benchmark Model Rank Results
image-generation-on-celeba-64x64DDPM++ (VP, FID) + ST#7FID: 1.9bits/dimension: 2.1
image-generation-on-celeba-64x64DDPM++ (VP, NLL) + ST#14FID: 2.9bits/dimension: 1.96
image-generation-on-celeba-64x64UNCSN++ (RVE) + ST#30bits/dimension: 1.97
image-generation-on-celeba-hq-256x256UNCSN++ (RVE) + ST#7FID: 7.16
image-generation-on-ffhq-256-x-256UDM (RVE) + ST#27FID: 5.54
image-generation-on-imagenet-32x32DDPM++ (VP, NLL) + ST#6FID: 8.42bpd: 3.85Inception score: 11.82
image-generation-on-lsun-bedroom-256-x-256UDM (RVE) + ST#7FID: 4.57
image-generation-on-stl-10UNCSN++ (RVE) + ST#2FID: 7.71Inception score: 13.43

Rank counts only results with a code link.