Trained on a 3080Ti with 443 sample images, 2 Vectors, Batch Size 2, 70,000 steps, with a custom learning rate schedule:
5e-4:70, 1e-3:210, 2e-3:350, 3e-3:560, 4e-3:770, 5e-3:1050, 6e-3:1400, 7e-3:2030, 8e-3:3640, 7e-3:4130, 6e-3:4620, 5e-3:5320, 4e-3:6160, 3e-3:7350, 2e-3:9240, 1e-3:11200, 9e-4:11620, 8e-4:12180, 7e-4:12880, 6e-4:13650, 5e-4:14700, 4e-4:16030, 3e-4:17920, 2e-4:21210, 1e-4:24640, 9e-5:25550, 8e-5:26670, 7e-5:27930, 6e-5:29540, 5e-5:31570, 4e-5:34300, 3e-5:38150, 2e-5:43960, 1e-5:48580, 9e-6:49630, 8e-6:50680, 7e-6:51940, 6e-6:53200, 5e-6:54740, 4e-6:56490, 3e-6:58590, 2e-6:61460, 1e-6:63700, 9e-7:64190, 8e-7:64750, 7e-7:65380, 6e-7:66150, 5e-7:66920, 4e-7:67970, 3e-7:69230, 2e-7
The schedule represents a short warm up, followed by a slightly tweaked exponential decay, which focuses on refining details over many steps.
