2023.05.27 updated
retrained model, removed image had rigging and other non-official outfit
also changed the trigger words
Recommend weight 0.8~1.0
Trigger words and supportive prompts
CHARACTER: yamato_\(kancolle\), 1girl,
※ Do not use character prompts if you want to get other outfit correctly.
ORIGINAL OUTFIT
(original_costume), headgear, detached_sleeves, red_skirt, pleated_skirt, black_thighhighs, bare_shoulders,
KAINI OUTFIT
(kaini_costume), headgear, detached_sleeves, long sleeves, white skirt, pleated skirt, black_highhighs, bare_shoulders, white coat, sleeveless coat, clothing cutout, coat on shoulders, long coat, metal collar, shoulder cutout, collar,
CASUAL OUTFIT
(spring_costume), alternate costume, official alternate costume, brown sweater, sweater, ribbed_sweater, long_sleeves, white_pants,
KIMONO OUTFIT
(newyear_costume), japanese_clothes, kimono, long_sleeves, wide_sleeves, obi, headgear, long hair, kimono, sash, floral print,
BIKINI OUTFIT
(summer_costume), swimsuit, bikini, white_bikini, sarong, official alternate costume, headgear, sandals, thigh_strap, navel, collarbone,
KAINI BIKINI
(summer2_costume), swimsuit, bikini, white_bikini, pink_bikini, sarong, bracelet, ribbon, hair_ribbon, see-through, navel, collarbone,
Training info
based model：nai anime full pruned
101 images for 4 repeats
original outfit 20 images
kaini outfit 18 images
casual outfit 17 images
kimono outfit 16 images
bikini outfit 15 images
kaini bikini 15 images
resolution : 768x768
rank=32, alpha=16
batch size = 2, epochs = 14
Optimizer：Lion
Learning rate = 3e-5, Unet learning rate = 1.5e-5, Text Encoder learning rate= 3e-5
======= old version =======
Lora of Yamato from kancolle with other alternative outfit
Recommend weight around 0.7
Recommend to increase the trigger words' weight to get better results
Trigger words and supportive prompts
CHARACTER
yamato_\(kancolle\), 1girl, brown_eyes, brown_hair, large_breasts, long_hair, hair_ornament, ponytail, cherry_blossoms, hair_flower
ORIGINAL OUTFIT
(yamato_original), headgear, detached_sleeves, red_skirt, pleated_skirt, thighhighs, bare_shoulders,
KAINI OUTFIT
(yamato_kaini), headgear, detached_sleeves, long_sleeves, white_skirt, pleated_skirt, thighhighs, bare_shoulders,
CASUAL OUTFIT
(yamato_casual), alternate_costume, sweater, ribbed_sweater, long_sleeves, sidelocks, white_pants,
KIMONO OUTFIT
(yamato_kimono), japanese_clothes, kimono, long_sleeves, wide_sleeves, obi, headgear,
BIKINI OUTFIT
(yamato_summer1), swimsuit, white_bikini, sarong, headgear, bare_shoulders, sidelocks, sandals, thigh_strap,
KAINI BIKINI
(yamato_summer2), alternate_costume, swimsuit, white_bikini, sarong, sidelocks, bracelet, ribbon, hair_ribbon, see-through,
May be difficult to generate kaini bikini outfit perfectly.
Training info
based model：Anything v4.5
216 images
original outfit 51 images for 4 repeats
kaini outfit 22 images for 8 repeats
casual outfit 22 images for 8 repeats
kimono outfit 18 images for 8 repeats
bikini outfit 20 images for 8 repeats
kaini bikini 16 images for 8 repeats
other outfit 67 images for 2 repeats
resolution : 768x768
rank=128, alpha=64
batch size = 2, epochs = 7
Optimizer：AdamW8bit
Learning rate = 1e-4, Unet learning rate = 1e-4, Text Encoder learning rate= 5e-5
Description
Network rank = 32, alpha = 16,
FAQ
Comments (9)
Using LoRA Block Weights to remove last ~5 blocks seems to help with the detail loss
MIDD:1,0,0,0,1,1,1,1,1,1,1,1,0,0,0,0,0
INS_MIDD:1,1,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0
thanks for advice!
XXXD you finished my job before me
Joke aside, since you have already made a dataset, and I've seen that your models have some problems like overfitted (most generated image are from same angle, and easily recognized to specific artist) and mixed elements like gloves and headgear from different object, undesired guns popping out here and there.
Here are some suggestions you might try out.
1. if the tag isn't misleading, don't remove it. If same tag is reference a visually different object like headgear, try to add \(name\) after it. e.g. headgear \(yamatokai2\). Delete tags to single trigger word do help with various concept charactor, but will cause the chara less detailed and have problems for a non-related outfits.
I've noticed that you removed a lot of tags, and trigger words to affect a lot, but it seems like the tag you kept has caused mixing elements.
Also, it's not recommended using character name as trigger word. The base model will mix it up by what it learnt.
2. shrink your dataset. 50 images are way too much for a single concept, especially this chara is very detailed, the quality of image is much more important than quantity. Suggest total images (images x repeats) up to 400. (Yours are 1122 now)
3. Dim rate should be under 64 for character LoRA. This is why your 128dim version is affecting the model's art style.
4. Try other learning optimizer, and use NovelAI leak (Anime-Full/Pruned) as base model for character lora. Adam8bit is bit out of date, and AnyV4 isn't a good base model for learning character.
Thanks your suggestions, I'm beginner and now im trying other optimizer and low Dim. I expect to remove other outfit images and lower the repeats.
Then, I didn't remove any tags, just didn't put them all on the prompts. Should I need to list them all?
@kazanesou434 Maybe you are using a high Threshold when tagging or you made some magic to make trigger word very effective (usually it won't be that strong if you didn't delete any tag), as far as I know, 0.7 is way too high for wd1.4 tagger to character. Personally I use 0.3 for all kind of training, otherwise it can easily lose detail. Also, the tag you SHOULD delete is proper noun like character names and too ambiguous tags like 'coat' or 'skirt'.
@4BEraser May be I didn't delete the concept tag of facial features or clothings, so the trigger words absorb all concept from the images.
@kazanesou434 yup, that definitely will cause clothing mixed up. It's hard to make multi-concept lora with close in tagger but very different looking outfits.
@4BEraser Got it, I will try it agian.
@kazanesou434 Oh, I forgot to tell you that you'd better separate the rigging illust dataset as a individual concept rather than combine into the original concept. It's definitely a bad idea when generating character only. That's why when your LoRA weight raise up, those undesired cannon popping up.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.