Trained on 52 pics (v1 was 34), so more flexible.

And a lower loss rate was achieved in the end, so more accurate.

```
steps: 30% 270/900 [06:30<15:11, 1.45s/it, loss=0.116]

steps: 40% 360/900 [08:42<13:04, 1.45s/it, loss=0.136]

steps: 50% 450/900 [10:56<10:56, 1.46s/it, loss=0.124]

steps: 60% 540/900 [13:08<08:45, 1.46s/it, loss=0.118]

steps: 70% 630/900 [15:21<06:34, 1.46s/it, loss=0.113]

steps: 80% 720/900 [17:33<04:23, 1.46s/it, loss=0.117]

steps: 90% 810/900 [19:46<02:11, 1.47s/it, loss=0.12]

steps: 100% 900/900 [22:05<00:00, 1.47s/it, loss=0.122]
```