/vt/ - Virtual Youtubers

>>38601835
Deepbooru interrogation gave me this from picrel:
1girl, animal ears, long hair, solo, breasts, swimsuit, purple eyes, tail, open mouth, bell, covered navel, looking at viewer, multicolored hair, one-piece swimsuit, smile, blush, very long hair, ahoge, tattoo, large breasts, black hair

Which was then appended to my quality tags and the embed name. Im still under the impression that its hard to quantify whether shuffling actually did anything worthwhile. I think in theory its good but in this case where I only go 1000 steps I only go over the dataset about 23 times or 23 times where the tags are shuffled. Compared to 10k training would have shuffled the tags 232 times to make the embed a lot more robust. But thats a 45 minute adventure just to train to 10k for me. 1k took like 5 minutes.

Ive done these types of tests with all my embeds Ive produced so far and ive seen a lot of thing. Some stuff ive come to associate with the quality of my embeds and others. I kind of wanted to explore even more training settings and see how the results change from that. There was a big training update a week ago or so that added a bunch of settings that I havnt barely touched yet. Im also experimenting to reduce my learn rate away from the default because a lot of my embeds gets overtrained and the difference between the 500 steps saves can be quite drastic.

Im doing all this to eventually get a more thorough embed training guide up to replace the one in the archive that has since the training update become outdated. and im still learning new things...