training: 

```
dataset = images-damian2
last 2 text encoder layers unfrozen
30 epochs at 2e-6, then 20 epochs at 1e-6 (step 240 on blue graph)
cond_dropout= default
```
first 30 epochs:
![](first_30_epochs.png)
resume:
![](resume.png)

sample:
![](sample.png)