| # Hakari Atsuko (Blue Archive) | |
| Note: her mask was not included in the training data. Stable Diffusion can't deal with them and a mixture of mask/no-mask Atsuko would fuck things up. | |
| ## Usage | |
| Use any or all of these tags to summon Atsuko: | |
| `1girl, halo, red eyes, pink hair, low twin braids` | |
| Her hair is sometimes tagged pink, sometimes purple, even on Danbooru. | |
| For her Arius outfit: | |
| `hood, white dress, hair bow, black gloves, white kneehighs, frilled legwear` | |
| For her leotard, just remove `white dress` and add `leotard` or `one-piece swimsuit` (WD unfortunately tagged it as a swimsuit more often than not). | |
| The AI likes giving her fairly prominent lips/lipstick for some reason despite her mouth being basically just a line in most of the training data. It feels a little excessive so I just neg-prompt `lips` and it looks closer to correct. | |
| You can add or ignore `atsuko, blue archive`; while they were in her captions, their effects aren't especially strong. | |
| Weights from 0.9 - 1.2 work well depending on variant and model. | |
| Variants: | |
| - v3 -- 3 epochs, faster learning rate; slightly underfit but less likely to appear overcooked | |
| - v5 -- 4 epochs, tweaked dataset, slower learning rate; does a great job with her outfit and halo but occasionally looks overcooked | |
| - v6 -- v3 with v5's dataset; largely a wash, different but not necessarily better or worse. | |
| Pick whichever, it doesn't really matter too much. Each one seems to have the chance of being the best for a given prompt so it's all RNG. | |
| ## Training | |
| *All parameters are provided in the accompanying JSON files.* | |
| - Trained on a set of 60 images initially provided by an /hdg/ anon, I pruned/added a few here and there. | |
| - Dataset included a mixture of SFW and NSFW. | |
| - Dataset was tagged with WD1.4 interrogator. Shuffling was enabled. | |
| - `atsuko, blue archive` were added to the start of each caption. keep_tokens=2 was set to prevent shuffling those tokens. | |
| - Three variations with different training params, check the JSON files if you care. | |