Training contents:

Black haired female characters. Common theme female characters with a disdainful , dangerous or cool edgy look.

Aesthetic reccomendation: Main priority illustrative / anime and or western cartoon focused images were used in training 50% of training data . Second priority semi-realistic 3DCG content 30% of training data . Third priority real photos of safe for work face and upper body shots 20% of training data.

Prompt recommendation:

Just use the built in abstract tool in TA (it among those buttons at the bottom of the promot text box). Use some tags + natural language.

I have adapted this script to insert the tags in the middle of the existing nat lang captions. E.g if the nat lang caption is 8 sentences long , then the tags will be inserted after 4sentences.

Note: captioning was a bit rushed and could only include dino tags , no nat lang descriptions. I used daily tensor pro tokens to train this at 60% discount , removing images from the Drills3D set I rarely made use of anyways when doing the previous lora , a throwing in images from Sankaku , from Pinterest , and some artstation stuff from Guweiz. Effective training equivalent to lora trained ysing 680 credits. Not bad for a rush job :)!

Body focus: Twintail braids , black hair , ponytails , glasses , shoulders , backsides , tanktops , casualwear t-shirt and short shorts , chokers , maid outfits , latex inspired clothing sleek or with straps , high legwear , hugo boss inspired female attire

Training data:

1) Guweiz (artist that make girls with long ponytails and tanktops from behind the shoulder viewes in hong kong esque cityscapes) 2) Wednesday Addams (illustrative works - no real photos) 3) safe for work pinterest photos as clothing references 4) Winter sisters from Marvels Gifted tv-show (disdainful look) 5) Sanudenuago (artist) + Erudetenis (artist) inpired works , including some simpsons character in similiar attire 6) Drills3D mixed 3DCG 7) As109 (Alice in red cheongsam , catgirls) 8) Stocking from PS&G 9) Monochrome manga on colored background of Eudetenis + Tatsumaki (pelvic curtain) from OPM + Shizuka (dark haired girl with glasses) from BP

Training method:

Each training frame is a 1024x1024 split into 3 column images. the border of traiming images are a 7px gray border. Learning rate 0.00035 cosine with repeats , 2 warmup steps , default text encoder learning rate.

At batch size 1 , 33 epochs x 3 repeats x 167 images = 16533 training steps , 16533 computation steps

Captioning were done using dino tagging (I didn't have time for nat lang captioning - hopefully its no big issue and lora will work using nat lang prompts). No keywords are used in this lora. Negatives are kept clear.