ResearchBenchmarksUnconditional Watermarked Generation on AFHQ 64x64 v2 (train)Unconditional Watermarked Generation on AFHQ 64x64 v2 (train)Follow2.1FID (50k Samples)MDM-proj1.91243.17874.4455.7113Oct 2, 2023Evaluation resultsMethodMethodLinksFID (50k Samples)FID* (50k Samples)MDM-projPrecision=56.9%Precision=56.9%2023.102.1—MDM-projPrecision=75.0%Precision=75.0%2023.102.12—MDM-dualPrecision=56.9%Precision=56.9%2023.102.232.21MDM-projPrecision=92.7%Precision=92.7%2023.102.3—MDM-dualPrecision=75.0%Precision=75.0%2023.102.862.32Prior watermarked diffusion model2023.104.32—MDM-dualPrecision=92.7%Precision=92.7%2023.106.793.05