| # Chidori Michiru (Blue Archive) | |
| ## Usage | |
| Use any or all of these tags to summon Koharu: | |
| `michiru, 1girl, halo, yellow eyes, grey hair, small breasts` | |
| Tag pruning means you don't need to prompt for much other than `michiru` to get a pretty good result -- eyes and hair are optional but can help correct occasional mistakes. | |
| The AI really tries to do her ninja `kuji-in` but unsurprisingly fucks it up more often than not, holding up the wrong number of fingers or just generating messed up hands. | |
| The tail is a bit iffy -- I tagged it as `tail` but prompting for that gets a cat tail half the time. Try `raccoon tail`. | |
| For her normal outfit: | |
| `school uniform, blue skirt, black pantyhose, floral print, black scarf, bridal gauntlets` | |
| I tagged images with `sarashi` even when it was only partially visible under her uniform. | |
| Unlike Koharu's LoRA I was a little more specific with clothing colors when tagging. Skirts are always tagged `blue skirt`, scarf is always `black scarf` etc. The hope was that it would make it possible to get her normal clothing in a different color, and to avoid overfitting `skirt` and `scarf` into the exact ones she normally wears. It sorta works but if you want a`red skirt` you have to emphasize it and negative prompt `blue skirt`. | |
| Weights from 0.8 - 1.05 should work well, it's perhaps slightly overtrained. Included both epoch 3 and epoch 4, epoch 3 might actually be a bit better. | |
| ## Training | |
| *All parameters are provided in the accompanying JSON files.* | |
| - Trained on a curated set of 88 images repeated 10 times. | |
| - Dataset included a mixture of SFW and NSFW. | |
| - This dataset was smaller than Koharu's, but I maintained the same number of steps. | |
| - Initially tagged with WD1.4, then performed heavy pruning and editing. | |
| - Removed as many inaccurate tags as possible | |
| - Made sure important traits were present and consitently described, and traits like `halo` were consistent with actual visibility | |
| - Pruned redundant tags and simplified outfits so that they were always tagged with the same handful of tags | |
| - Added camera angles and image composition hints | |
| - Added a few facial expressions | |
| - Different learning rate than usual. | |
| - 5e-5 text encoder (same as Koharu's, but typically 1e-5 ~ 2e-5) | |
| - 3e-4 UNet (typically one order of magnitude faster than text) | |
| - Trained without VAE. | |