Duisburg (Azur Lane) [Character] (NoobAI)版本v4.0 (ID: 1270783)

Character model based on 10 15 31 43 images of Duisburg from gacha game Azur Lane. Very good reproducibility of the character and outfits now since v3.

Usage

Basic

Works best at a strength of 1.0 if you want accurate base outfits, turn it down to 0.8 if you are getting outfit bleed.
Base trigger is duisburg

Outfits

Outfit components are tagged separately so you can choose what to include.

• Base skin:
  black sweater, turtleneck, black jacket, red belt, open jacket, cleavage cutout, underboob, black thighhighs, thigh strap, gloves, manjuu, black boots

• Endurance Training (Gone Wrong) (not very accurately reproducible on v1.0):
  rabbit ears, hairclip, choker, black lingerie, lace-topped thighhighs, lace-topped panties, lace bra, between breasts, detached sleeves

Issues

After much frustration at strange outputs on some seeds with versions 1 through 3, I think I finally figured out what's been happening.

The most consistent things coming through were a Blue Archive-esque halo, and black hair, sometimes with a pink underside:

After scrolling through some BA characters, I finally found the culprit: Kazusa

My guess is that Blue Archive data is so massively overrepresented on Danbooru that it skews the output of NoobAI towards those characters, i.e. it saw Duisburg's pink-ish eyes, short hair and red streak and thought "that must be Kazusa".

In v4.0 I have attempted to mitigate this by reducing the training rate a little and training for more steps, 'overtraining' a little bit to hopefully make sure the model's base info is overriden by Duisburg's appearance. It seems to work a lot more consistently now, but I did still have an ~1/40 failure rate where the output was infected by what I'm calling the "Blue Archive Plague".

Extra

In the interest of openness and community-helping, I have uploaded the dataset along with the model, this includes: images, captions, and the training json (for kohya_ss).

描述:

Wasn't really expecting to do another version, but after seeing repeated issues in v1-3 I think I have finally discovered what is 'wrong' with this character on NoobAI, the plague of Blue Archive (specifically 'Kazusa'). See issues section of main description for more info.

• trained on 43 (+12) images on NoobAI ε-pred 1.0

• modified training parameters to train for a longer duration at a lower rate

• much less character bleed of Kazusa (blue archive)

训练词语: duisburg,(see description)