lora-training / chise /README.md
khanon's picture
initial commit
7529c6f
|
raw
history blame
1.99 kB

Waraku Chise (Blue Archive)

Came out pretty well I think. Smaller dataset than Mari, but otherwise very similar settings.

Usage

Use any or all of these tags to summon Chise: chise, halo, red eyes, blue hair Hair and eyes are mostly optional if you describe a bit of her outfit as well. She naturally likes to make her :o expression because most art features her doing that. However I also included images tagged with other expressions. Use open mouth, closed mouth, and parted lips as necessary to get her to make whatever expressions you want.

For her normal outfit (add as many as necessary): braid, japanese clothes, detached sleeves, obi, tabi, geta, sailor collar, blue bow Her shoes are weird but they're tagged geta and the socks as tabi.

For her swimsuit outfit (add as many as necessary): side ponytail, swimsuit, striped bikini, see-through, sailor collar, side-tie bikini bottom You can also add pointed ears, which are usually only visible in her swimsuit outfit.

Weight 1 works fine. Also included epoch 6 in case you find the last epoch to be a bit stubborn with her outfits.

Training

All parameters are provided in the accompanying JSON files.

  • Trained on a set of 119 images split by outfit, repeated 10 times (swimsuit) or 14 times (uniform) for 7 epochs (119 images * 10 or 14 repeats / 3 batch size * 7 epochs = 3178 steps)
    • Dataset included a mixture of SFW and NSFW.
  • Initially tagged with WD1.4 VITv2 model, then performed heavy pruning and editing.
    • Pruned implicit (oni horns) or 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
  • Used network_dimension 128 (same as usual) and network_alpha 64 (new)
    • This relies on the new alpha
  • Trained without VAE.