v3.0 & v3.1
A LoRA based on 742 Cinderella Girls game art, recognizing the 9 characters in the U149 anime. A small test for all-character models.
Replaced Japanese character tags. Removed Hadamard Product dependency. Reduced file size. Added deremas card concept.
Usage
The style itself doesn't have a trigger word. For the characters, below are the character tags and respective top 20 related tags.
Sasaki Chie,
short hair, black hair, black eyes, smile, 1girl, solo, hair ornament, hairclip, open mouth, rabbit hair ornament, looking at viewer, blush, skirt, bangs, dress, holding, card, bow, hat, braid
Akagi Miria,
short hair, black hair, smile, brown eyes, two side up, 1girl, open mouth, solo, looking at viewer, blush, hair ornament, :d, skirt, brown hair, gloves, one eye closed, jewelry, star, dress, closed eyes
Ryuzaki Kaoru,
hair ornament, short hair, hairclip, brown hair, smile, 1girl, open mouth, solo, yellow eyes, looking at viewer, shorts, orange hair, hat, closed eyes, card, character name, sun symbol, jewelry, flower, food
Tachibana Arisu,
brown eyes, brown hair, long hair, 1girl, solo, bow, smile, hair bow, dress, looking at viewer, open mouth, blush, ribbon, black hair, flower, food, frills, hair ornament, fruit, white background
Ichihara Nina,
long hair, smile, brown hair, open mouth, 1girl, solo, brown eyes, blush, cosplay, animal costume, kigurumi, bangs, :d, looking at viewer, bow, blunt bangs, closed eyes, hat, striped, gloves
Matoba Risa,
black hair, long hair, yellow eyes, twintails, 1girl, solo, ribbon, hair ribbon, smile, jewelry, necklace, open mouth, skirt, looking at viewer, blush, midriff, bangs, one eye closed, belt, gloves
Yuki Haru,
orange hair, 1girl, solo, smile, long hair, open mouth, shorts, brown eyes, looking at viewer, gloves, midriff, holding, navel, blush, brown hair, hat, shirt, sky, fingerless gloves, white background
Sakurai Momoka,
blonde hair, green eyes, short hair, smile, 1girl, solo, looking at viewer, dress, open mouth, flower, wavy hair, hairband, frills, blush, hair ornament, bow, gloves, jewelry, rose, hat
Koga Koharu,
short hair, 1girl, solo, brown hair, smile, brown eyes, flower, open mouth, dress, :d, bow, card, character name, hairband, pink background, looking at viewer, hair bow, gloves, star, blush
Training info
Dataset:
- 129 framed Cinderella Girls cards from アイマス シンデレラガールズ カードギャラリー (no frameless version available)
- 273 frameless Cinderella Girls cards from アイマス シンデレラガールズ カードギャラリー
- 174 Starlight Stage cards ("spreads") from Starlight DB
- 166 Starlight Stage character cutouts ("sprites") from Starlight DB
Tags generated with SmilingWolf/wd-v1-4-convnextv2-tagger-v2 with a threshold of 0.35. No tag cleaning. Characters roughly balanced for each image type so that #max / #min <= 2. Denoised with RealESRGAN_x4plus_anime_6B.
v3.0 trained on resolution 512 for 63 epochs. v3.1 continued from v3.0 and trained on resolution 768 for 15 epochs. A custom SGDR lr scheduler was used.
Training cost: ~5.5 T4-hours
Data distribution:
character framed frameless sprite spread total
1 Sasaki Chie 21 30 24 22 97
2 Akagi Miria 13 37 20 18 88
3 Ryuzaki Kaoru 16 34 20 18 88
4 Tachibana Arisu 7 34 24 22 87
5 Ichihara Nina 6 38 20 20 84
6 Matoba Risa 19 22 20 18 79
7 Yuki Haru 13 26 20 18 77
8 Sakurai Momoka 6 36 10 24 76
9 Koga Koharu 28 16 8 14 66
Training config:
[model_arguments]
v2 = false
v_parameterization = false
pretrained_model_name_or_path = "../animefull-final-pruned.ckpt"
[additional_network_arguments]
no_metadata = false
network_module = "networks.lora"
network_dim = 8
network_alpha = 1
[optimizer_arguments]
#optimizer_type = "DAdaptation"
optimizer_type = "AdamW8bit"
learning_rate = 3e-4
max_grad_norm = 1
lr_scheduler = "custom"
[training_arguments]
save_precision = "fp16"
save_every_n_epochs = 1
train_batch_size = 2
max_token_length = 225
mem_eff_attn = false
xformers = true
max_train_epochs = 9999
max_data_loader_n_workers = 8
persistent_data_loader_workers = true
gradient_checkpointing = false
gradient_accumulation_steps = 1
mixed_precision = "fp16"
clip_skip = 2
lowram = true
cache_latents = true
min_snr_gamma = 5
[saving_arguments]
save_model_as = "safetensors"