v2 Increased images from 146 to 217. Replaced a few images with higher quality. Increased network_dim from 32 to 128. Increased training resolution from 512 to 576. Increased offset_noise from 0.0 to 0.0025. Decreased epochs from 15 to 10.
Seems a little cooked, so you may have to decrease weight to 0.8 0.9.
Does any one know what training setting causes this? Is it the network_alpha? I set mine to 64, half of network_dim:
Description
pics: 217
blip: 20 - 100
resolution: 576
num_repeats: 10
num_epochs: 10
network_dim: 128
network_alpha: 64
min_snr_gamma: 0.0025
steps: 10% 364/3640 [11:50<1:46:37, 1.95s/it, loss=0.0842]
steps: 20% 728/3640 [23:51<1:35:24, 1.97s/it, loss=0.0852]
steps: 30% 1092/3640 [35:51<1:23:40, 1.97s/it, loss=0.0836]
steps: 40% 1456/3640 [47:52<1:11:48, 1.97s/it, loss=0.0781]
steps: 50% 1820/3640 [59:54<59:54, 1.97s/it, loss=0.0814]
steps: 60% 2197/3640 [1:12:33<47:39, 1.98s/it, loss=0.0791]
crashed: out of memory (15 GB card)
steps: 25% 364/1456 [11:52<35:36, 1.96s/it, loss=0.0788]
steps: 50% 728/1456 [23:55<23:55, 1.97s/it, loss=0.0788]
steps: 75% 1092/1456 [35:57<11:59, 1.98s/it, loss=0.076]
steps: 100% 1456/1456 [47:58<00:00, 1.98s/it, loss=0.0766]
FAQ
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.