Trained on 150K images from Rule34.xxx for 5 epochs.
Second epoch commentary
I yeeted about 30K images and then added 20K back from good artists, and lowered the LR significantly to reflect my batch size, I think it's improved significantly but I didn't save any prompts/settings from epoch0 to compare it with.
Third epoch commentary
Third epoch, again a pretty nice improvement over the last. Dataset grew by ~3K images, with 5K removed and 8K added.
Fourth epoch commentary
r34 e3 (fourth epoch). Same dataset, prepended "by" to artist tags, artist recognition has improved a little bit (still needs to get better), character recognition has improved quite a bit.
There is no commentary for epochs 0 and 4. (Epochs were counted from zero.)
