# Iochi Mari (Blue Archive)
I used a bigger dataset than normal since Mari has a lot of good art (253 images).

This was a tricky LoRA to train since she has two very distinct outfits (and her gym outfit has as much art without her jacket as with). The LoRA recalls both outfits, and even her swimsuit, but you may need to negative prompt elements of whichever outfit you don't want.

## Usage
Use any or all of these tags to summon Mari:
`mari, halo, 1girl, blue eyes`
`cat ears` was only tagged on images where her cowl was not present, in which case `animal ear headwear` is the tag you want.
Hair and eyes are optional.

For her normal outfit (you don't usually need all of these):
`single braid, animal ear headwear, nun, habit, puffy long sleeves, blue neckerchief, white sailor collar`
Her shoes were tagged `mary janes`.
The sleeves of her outfit leak into other outfits/nudity so you may need to negative prompt them (`puffy long sleeves`)

For her gym uniform outfit (you don't usually need all of these):
`low ponytail, sportswear, track jacket, shorts, gym shirt, ponytail, id card, bottle`
Images with her jacket visible were tagged `track jacket`. If the jacket was open, I added `open jacket`.
Images with her gym shirt visible (so not those where her jacket was closed) were tagged `gym shirt`.  Adding `white` can help.
`wet, wet clothes` was common in the trianing data and works well.

For her swimsuit (only a handful of images in the training data, so use emphasis):
`black one-piece swimsuit, frilled swimsuit, frills, paw print`
Negative prompt: `competition swimsuit`

Included three epochs.  I did way more steps than normal in pursuit of perfect halo gens, 
- epoch 7 (3542 steps) = nails her expressions and outfits pretty well but outfit details can leak over. best halo but still not great
- epoch 6 (~3000 steps) = pretty good balance between the others
- epoch 4 (~2000 steps) = less overfit on outfits but halo and expression consistency is not as good

Weight 1 works fine for the most part.  Epoch 7 may be somewhat stubborn/overfit on some parts of her outfit, in which case you may have better luck dropping down to epoch 6 or 4.


## Training
*All parameters are provided in the accompanying JSON files.*
- Trained on a set of 253 images repeated 6 times, 7 epochs (253 images * 6 repeats / 3 batch size * 7 epochs = 3542 steps)
  - Dataset included a mixture of SFW and NSFW.
  - Same params as Koharu, but with almost double the amount of steps, since the dataset was large and I wasn't initially satisfied with halo/facial expression consistency
- Initially tagged with WD1.4, then performed heavy pruning and editing.
  - Pruned redundant tags and simplified outfits so that they were always tagged with the same handful of tags
  - Made sure important traits were present and consitently described, and traits like `halo` were consistent with actual visibility
  - Added many facial expression, camera angle, and image composition hints
- Made a hard-coded tweak to bmaltais's LoRA GUI script.
  - Previously increasing the `resolution` values (512 >> 768 >> 832) would not increase `max_bucket_reso`, so buckets were limited to 1024 (the default) in either dimension.
  - Per the kohya_ss README, when increasing `resolution` to 768, it is recommended to increase `max_bucket_reso` to 1280 to keep the bucket aspect ratios the same as 512 resolution
  - I adjusted the script to set `max_bucket_reso` based on `resolution`, using the formula:
    - `max_bucket_reso = resolution * (1 + (512 / resolution))` which causes bucket aspect ratios to be the same as 512 for any given resolution.
  - In practice, this only affected a very small number of very tall or very wide images so I don't think it would have made much of a difference
- Trained without VAE.