2 vectors. Vectors were trained individually and then merged. Probably some potential for improvement by training a third or more vectors but I'm pretty satisfied with the results here. The main area that could be improved is face distortion with more distant shots, although using "restore faces" or img2img at a higher resolution can often solve any weirdness with that.
I will likely work on a SD 2.1 based version soon and upload that as well.
Batch Size 24, 20000 steps for each vector, trained on a 3080Ti
Custom learning rate schedule:
8e-3:40, 7e-3:160, 6e-3:320, 5e-3:520, 4e-3:780, 3e-3:1140, 2e-3:1720, 1e-3:2300, 9e-4:2440, 8e-4:2620, 7e-4:2820, 6e-4:3060, 5e-4:3360, 4e-4:3760, 3e-4:4320, 2e-4:5320, 1e-4:6340, 9e-5:6620, 8e-5:6960, 7e-5:7340, 6e-5:7820, 5e-5:8440, 4e-5:9260, 3e-5:10420, 2e-5:12160, 1e-5:13560, 9e-6:13860, 8e-6:14200, 7e-6:14560, 6e-6:14960, 5e-6:15400, 4e-6:15920, 3e-6:16560, 2e-6:17420, 1e-6:18100, 9e-7:18260, 8e-7:18440, 7e-7:18620, 6e-7:18840, 5e-7:19080, 4e-7:19380, 3e-7:19760, 2e-7
