Lora model of Mizuki Yukikaze from Taimanin games. Most previews done with weight 1, might have been some with weight 0.9.
The first gacha game I played was Taimanin Kessen Arena and this was the best girl :D I actually planned to just do normal version and adult version but somehow I ended up doing way more, surely most outfits I ever put together. They are not perfectly accurate but "about right" and sometimes using some negative prompting can prevent blending together. The model is a bit overfit, but highres fix helps with quality (I don't think it overfits poses much). I might do a fix sometime in future because previous version that I made had no overfitting so I know I went wrong somewhere at the last turn >_< I want to move on to other things now through.
Prompts and tips:
Usual outfit - mizuki yukikaze, hair ribbon, one-piece tan, completely nude - putting white one-piece swimsuit in negative can help because tanlines look a lot like swimsuit and model tends to add outline to it.
Taimanin suit - mizuki yukikaze, hair ribbon, black leotard, taimanin suit, one-piece tan, elbow gloves, fur trim, fingerless gloves, frilled leotard, thigh boots - fur trim can lower weight or remove it if it bothers you. Adding red gloves, red thighhighs, cleavage to negatives can help with accuracy.
Transparent swimsuit - mizuki yukikaze, hair ribbon, white one-piece swimsuit, one-piece tan, see-through, covered nipples - lowered Lora weight helps here I think.
Bikini - mizuki yukikaze, hair ribbon, black choker, one-piece tan, micro bikini, white bikini, white gloves, bridal gauntlets, see-through, high heels
Armored bodysuit - mizuki yukikaze, hair ribbon, black bodysuit, fishnets, pauldrons, gauntlets, armored boots
Black dress-like outfit - mizuki yukikaze, hair ribbon, frilled collar, brooch, one-piece tan, breastless clothes, detached sleeves, corset, frills, heart cutout, micro panties, bridal garter, high heels
Jacket outfit - mizuki yukikaze, hair ribbon, cropped jacket, puffy sleeves, see-through leotard, see-through, one-piece tan, elbow gloves, grey gloves, thigh boots
Red outfit - mizuki yukikaze, hair ribbon, red leotard, elbow gloves, red gloves, showgirl skirt, highleg panties, red thighhighs, thigh boots - adding something like black bodysuit or covered navel to negative can help.
School uniform - mizuki yukikaze, hair ribbon, school uniform, green shirt, serafuku, white neckerchief, long sleeves, pleated skirt, black skirt, black thighhighs, loafers
School uniform (another one) - mizuki yukikaze, hair ribbon, school uniform, white shirt, short sleeves, blue necktie, pleated skirt, blue skirt, white thighhighs, loafers
School uniform (why she have three??) - mizuki yukikaze, hair ribbon, school uniform, white shirt, brown vest, long sleeves, red necktie, pleated skirt, black skirt, white thighhighs, loafers
Adult version - mizuki yukikaze, short hair, taimanin suit, black leotard, fishnets, elbow gloves, vambraces, armored skirt, thigh boots
Adult evening dress - mizuki yukikaze, short hair, hair flower, white capelet, black choker, earrings, necklace, black dress, plunging neckline, elbow gloves, brown gloves, navel cutout, thong, bracelet - lower Lora weight can help here cause this one is lower quality
Adult swimsuit - mizuki yukikaze, short hair, casual one-piece swimsuit, white choker, see-through, halterneck - I actually tagged this one in two ways so you can try white one-piece swimsuit instead or both, depending on picture.
Description
FAQ
Comments (17)
你做的好啊
Thank you so much!
Just curious, what training parameters did you use? I tried inspecting with Additional Networks but it wasn't able to read the metadata
You wouldn't be able to see it because the model is a merge between two versions. The bigger part parameters look like this:
{
"ss_adaptive_noise_scale": "None",
"ss_batch_size_per_device": "1",
"ss_bucket_info": "{\"buckets\": {\"0\": {\"resolution\": [256, 640], \"count\": 2}, \"1\": {\"resolution\": [256, 768], \"count\": 22}, \"2\": {\"resolution\": [320, 576], \"count\": 31}, \"3\": {\"resolution\": [320, 640], \"count\": 4}, \"4\": {\"resolution\": [320, 704], \"count\": 17}, \"5\": {\"resolution\": [320, 768], \"count\": 5}, \"6\": {\"resolution\": [384, 448], \"count\": 12}, \"7\": {\"resolution\": [384, 512], \"count\": 180}, \"8\": {\"resolution\": [384, 576], \"count\": 286}, \"9\": {\"resolution\": [384, 640], \"count\": 34}, \"10\": {\"resolution\": [448, 448], \"count\": 41}, \"11\": {\"resolution\": [448, 512], \"count\": 163}, \"12\": {\"resolution\": [448, 576], \"count\": 28}, \"13\": {\"resolution\": [512, 384], \"count\": 320}, \"14\": {\"resolution\": [512, 448], \"count\": 30}, \"15\": {\"resolution\": [512, 512], \"count\": 49}, \"16\": {\"resolution\": [576, 384], \"count\": 14}, \"17\": {\"resolution\": [576, 448], \"count\": 25}, \"18\": {\"resolution\": [640, 384], \"count\": 44}, \"19\": {\"resolution\": [704, 320], \"count\": 13}}, \"mean_img_ar_error\": 0.027568858431215225}",
"ss_bucket_no_upscale": "True",
"ss_cache_latents": "True",
"ss_caption_dropout_every_n_epochs": "0",
"ss_caption_dropout_rate": "0.0",
"ss_caption_tag_dropout_rate": "0.0",
"ss_clip_skip": "2",
"ss_color_aug": "False",
"ss_enable_bucket": "True",
"ss_epoch": "12",
"ss_face_crop_aug_range": "None",
"ss_flip_aug": "True",
"ss_full_fp16": "False",
"ss_gradient_accumulation_steps": "1",
"ss_gradient_checkpointing": "False",
"ss_keep_tokens": "1",
"ss_learning_rate": "1.0",
"ss_lowram": "False",
"ss_lr_scheduler": "constant",
"ss_lr_warmup_steps": "0",
"ss_max_bucket_reso": "1024",
"ss_max_grad_norm": "1.0",
"ss_max_token_length": "225",
"ss_max_train_steps": "19800",
"ss_min_bucket_reso": "256",
"ss_min_snr_gamma": "5.0",
"ss_mixed_precision": "bf16",
"ss_multires_noise_discount": "0.3",
"ss_multires_noise_iterations": "None",
"ss_network_alpha": "16.0",
"ss_network_dim": "64",
"ss_network_dropout": "None",
"ss_network_module": "networks.lora",
"ss_new_sd_model_hash": "89d59c3dde4c56c6d5c41da34cc55ce479d93b4007046980934b14db71bdb2a8",
"ss_noise_offset": "None",
"ss_num_batches_per_epoch": "1320",
"ss_num_epochs": "15",
"ss_num_reg_images": "0",
"ss_num_train_images": "1320",
"ss_optimizer": "dadaptation.dadapt_adam.DAdaptAdam(decouple=True,weight_decay=0.01,betas=(0.9, 0.99))",
"ss_output_name": "MizukiYukikazeV1",
"ss_prior_loss_weight": "1.0",
"ss_random_crop": "False",
"ss_reg_dataset_dirs": "{}",
"ss_resolution": "(512, 512)",
"ss_scale_weight_norms": "None",
"ss_sd_model_hash": "925997e9",
"ss_sd_model_name": "nai.ckpt",
"ss_sd_scripts_commit_hash": "fa8fbe1ac1cf2a1356a56a596521118ccb8cdeb3",
"ss_seed": "1234",
"ss_session_id": "588935058",
"ss_shuffle_caption": "True",
"ss_steps": "15840",
"ss_text_encoder_lr": "1.0",
"ss_total_batch_size": "1",
"ss_training_comment": "None",
"ss_training_finished_at": "1686592729.9413836",
"ss_training_started_at": "1686580859.474874",
"ss_unet_lr": "1.0",
"ss_v2": "False"
}
@Deto15 is it possible for you to share the dataset? just curious how you tag and managa all of this dataset folder. I tried to train multi concept lora like this but still fail although i already follow all the step from hollowstrawberry guide
@chronus I'm worried about posting full dataset but I packed few examples https://drive.google.com/file/d/1iEM4it56QTPNrJWOv1AUsc6rcdo1edNz/view?usp=sharing
I don't really have an established worflow through and try out many things. Here for example I used weight captions a little bit for first time.
I skipped through the guide you mentioned and I'm not sure if it mentions that you can split your dataset into few folders with different numbers in front of them so that some images get looked at more in epoch than the other. If you are training model with few outfits or few concepts then that's often a must. I had separate folder for each outfit here.
@Deto15 Thank you for giving me your dataset example, although I still don't know why my lora always fails. I already tag everything inside my image and remove tags that describe the character or clothes I want to train. But my lora just refuses to learn that my character is wearing a long purple slit skirt; they always give me a short skirt with a random colour when I use it. I already make sure to give a full body image, so I guess I will just give up for now.
as for training with hollowstrawberry guide or any colab, you can train one folder more than the other or use different number of token. You can also change the resolution and some other thing but i forget how. Here is the example setting that you can use.
[[datasets.subsets]]
image_dir = "/content/LoRA/train_data/azcasual"
num_repeats = 2
keep_tokens = 1
[[datasets.subsets]]
image_dir = "/content/LoRA/train_data/azarmor"
num_repeats = 1
keep_tokens = 2
@chronus If you are training outfits, have you tried simply using captions that model already knows like I usually do? I usually don't train new words to do that. For what you describes something like "purple dress, side slit" might work. Also it's possible that the amount of training you do on that outfit is still lacking. Even if you train two times more, if you have ten times less images it will still be a problem. So guides will tell you that you should balance so that each concept is equal with images number * repeats number.
@Deto15 i guess i will try using "purple dress, side slit" tag to see it improve.
as for amount of training, i already use 75 image with 2 - 4 repeat and train for 40 epoch or more untill it's overbaked, but the skirt is just straight up ignored like it doesnt even exist. Everything else like her hair, accesories, sleeves and shirt is okay. It's like it only train upper half of her body lol. How is that even possible? more than half of my image already include full body image so i'm 100% clueless now (Ｔ▽Ｔ)
will try to adjust the caption to see if there is any improvement.
@chronus Hope the tags will help. This is just a wild guess but maybe you're not using buckets? Because technically I think your images could be getting cropped if you have various resolutions and no buckets enabled (through I doubt result would be this specific).
I usually train through Kohya GUI so I don't have the dataset setup file like you posted, but I had something like that when I was training on Colab. You can see example dataset config here with buckets enabled https://drive.google.com/file/d/1snBV7OofvDOAvTrZrFOKSFSibb6lVPf8/view?usp=sharing
@Deto15 after i retrain my lora, it seems like leaving slit skirt tag in all my image instead of removing them give me better result. Weird since most guide i read tell me to remove any tag that describe my character. I guess it's because AI just doesn't understand how "slit skirt" works. You can try create image with "slit skirt" tag and they sometimes will give you pelvic curtain clothes.
I really don't understand how AI think or how this lora train works.
@chronus I suggested you "purple dress, side slit" because I know models based on Nai should be able to know those things. You went with "slit skirt" which they indeed don't know I think. Also if you asked me I would say "slit skirt" doesn't describe character, it describes clothing of character. You can see in prompting on my model that I don't identify clothing with character.
Some people do through. It's a difference in approach.
Also you say you added it to all your images but is it actually visible in all your images? You don't tag what you can't see.
thank you, truly appreciate the work you put into this, this is amazing, she's easily in my top 3. here's hoping someone does one for Maika
I don't even remember her ^_^; I might do Sakura at some point since she was my second favorite.
@Deto15 i don't blame you, i think she's a bit obscure but i just recently discovered her cause of Action Taimanin and she's probably my top 2, Ingrid being my top 1 :D
this one is good, please do more taimanin characters
If only we could port her to pony XL!!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.