Melty (Shield Hero)

[wild pagoda]

huggingface: https://huggingface.co/onusai/LoRAs#sh-melty
civitai: https://civitai.com/models/6355/melty-shield-hero

[crude sigil]

I made Raph, you made Melty, which one of us is making Filo?

[wild pagoda]

lol nice i didnt know you already made raph, im not planning to make anymore characters from shield hero so it'll up to you

[crude sigil]

Raph is somewhere on the embeddings page.

And fair enuff, if I get bored enough I'll do Filo at some point. Would much rather do Sadeena except she has next to no art.

[wild pagoda]

rip yea sadeena only has like 3 decent images on gelbooru. melty was also pretty rough to train bc she has very few fully clothed sfw images, ended up using like 50+ screenshots from anime

[crude sigil]

Ooooooooooooooooooooooof.

[sharp helm]

can you provide some advice for how you trained this model - I want to do malty as well

[civic crag]

time to lewd her into oblivion

[civic crag]

didnt know people liked her that much

[wild pagoda]

heres some of the parameters from the config i used:

image_count: 104
n_repeats: 3
num_train_images: 312 (including repeats)

batch_size: 8
epochs: 4
network_dim: 128
network_alpha: 32

learning_rate: 1e-05
unet_lr: 0.001
text_encoder_lr: 5e-05
lr_scheduler: cosine

shuffle_caption: True
keep_tokens: 1
flip_aug: True
color_aug: False
random_crop: False

clip_skip: None (probably means 1)
max_token_length: 225
sd_model_name: animefull-final-pruned.ckpt

i used Grabber to get about 50 pictures of melty from gelbooru but there was very little fully clothed sfw images, so the resulting lora was very nsfw and didnt capture her main outfit. i needed a lot of images of her outfit so i decided to try using anime screenshots

heres the process i used for screenshots:

1. i took about 56 unique screenshots of melty from one episode (she appeared a lot in it) while watching the episode in fullscreen (to get higher res screenshots)

2. then i divided them into 2 folders. the first folder contains all images that have a dimension smaller than 512 pixels, the second folder contains all images where both dimensions are above 512

3. used the auto1 webui to batch upscale the first folder at 2x using R-ESRGAN 4x+ Anime6B. then batch upscaled the second folder at 1x (made the images a bit sharper)

4. (optional?) i did a lot of iterations so im not sure if the lora i uploaded was before or after i did this, but this may be useful: I used the ABG_extension to remove backgrounds from all the screenshots leaving only melty. i had to edit some images manually to fully remove the parts ABG missed. then i asked chatgpt to make me a python script that removes the alpha channel from all the images and replaces it with white

5. used the kohya_ss webui to batch tag all the images using wd1.4, then used a script to add melty, to the start of all the .txt files

think thats about it, lmk if u have any questions

[wild pagoda]

tbh idk if many people do

i put more effort into this lora than id like to admit xD