Ryuuou no Oshigoto! // (complete pack Illustrious, PDXL and SD1.5) - hinatsuru ai SD1.5

NSFW

Support My Work:

Your support helps me create more amazing content and keeps this passion project alive. Every contribution, no matter how small, makes a big difference. Thank you for being awesome!

old text:

Add a ❤️ and follow to support my future works.

Every character in a different version of this LoRa:

Additional info about choised LoRa in "About this version block"

//why i change format:

it's bad if your LoRa's don't sort in automatic1111 lora list by title(title need to be first);
it's bad if your don't see trigger word in name of model(and civitai crop's long names, but for version it have additional characters);
it's not so good to made 20 LoRa's for each title on Civitai, just look like spam.
multicharacter LoRas harder to control->it's bad.

// result is format <title name>_<character trigger word> // model can contain additional TW, like "Pajamas", but this is not necessary//

train models isn't easy and you can support me here:

❤️Boosty Donation link

LoRa trained on this model https://civarchive.com/models/49608/randomizer89merge

Description

try to made her more controllable and sharp, additional TW list:

main dress style

pajamas

white berets

backpack

use it as positives to get this items more stable.

use it as negatives if you want made other dress without color shifting/leakage.

autistic eyes, horror eyes, enthusiastic eyes - use it's as negative if you want more clearer eyes.

if it's hard to control for you - use old one, current one need more hard tagging, because it's little overtrained.

if your model not anime use old one too, this one is overtrained, and you can't get good results on realistic models.

FAQ

Comments (6)

tekkanin0925yaho3018May 3, 2023

CivitAI

Question: How many images are you using to learn this?

randomizer89

Author

May 3, 2023· 1 reaction

91 for Hinatsuru Ai, 47 for Charlotte. // not many

https://civitai.com/models/46095/astarotte-astarotte this one is 571 -> better to do 50-100 high quality, than 600 all you can get.

tekkanin0925yaho3018May 3, 2023

@randomizer89 It was helpful Thank you!

randomizer89

Author

May 3, 2023· 1 reaction

@tekkanin0925yahoocojp186 really default 100 steps*image count make good results if your images in dataset is good quality and you made character(for styles it's different).

// i made more steps on this just for sharper results on main character, side characters not need it, hires fix anyway made sharp//

Anarky9May 10, 2023

@randomizer89 any time I try to train with 100 steps per image it always comes out very overtrained and distorted.. what learning rates do you use? Thanks!

randomizer89

Author

May 10, 2023· 2 reactions

@kalencunningham kiyotaki keika:

"61 image, 110 steps image"

cmd:
accelerate launch --num_cpu_threads_per_process=2 "train_network.py" --pretrained_model_name_or_path="F:/gamedev/SDwebui/webui/models/Stable-diffusion/7thHentaiHeaven_randomizer89fp16pruned.safetensors" --train_data_dir="F:/LoRa/TrainData/Ryuuou no Oshigoto!/Keika Kiyotaki" --resolution=512,512 --output_dir="C:\Users\Admin\Desktop" --network_alpha="128" --save_model_as=safetensors --network_module=networks.lora --text_encoder_lr=4e-05 --unet_lr=8e-05 --network_dim=62 --output_name="Ryuuou no Oshigoto!_kiyotaki keika" --lr_scheduler_num_cycles="1" --learning_rate="8e-05" --lr_scheduler="polynomial" --lr_warmup_steps="671" --train_batch_size="1" --max_train_steps="6710" --save_every_n_epochs="1" --mixed_precision="fp16" --save_precision="fp16" --seed="1" --optimizer_type="AdamW" --max_data_loader_n_workers="1" --max_token_length=150 --clip_skip=2 --bucket_reso_steps=32 --xformers --bucket_no_upscale --random_crop --sample_sampler=euler_a --sample_prompts="C:\Users\Admin\Desktop\sample\prompt.txt" --sample_every_n_steps="300"

i just copypast it from console(don't have others because of restarts, i change it time to time by small values to find optimal), you can watch result(here) and command, maybe it "overtrained and distorted" for you

LORA

SD 1.5

by randomizer89

Download (Beta) View on CivitAI