[LoRA] Shibuya Kanon (Love Live! Superstar!!)版本v2 (ID: 82716)

Update: SDXL 1.0 version here, with additional ability to generate Kanon with glasses and training outfit.

Trained using around 150 anime screenshots from season 1 and season 2. 20-30 images are chosen from sideway angles (booru tag from side). This is important, as previous models I trained for Kanon as well as other characters lacked the ability to generate sideway images.

Activation tag: shibuya kanon

To generate school uniform: student uniform, gray dress, red ribbon, neck ribbon, pinafore dress, white shirt

To generate glasses: glasses. For v2, include bangs, hair between eyes as well (the model will generate her wearing glasses without bangs if not prompted)

Notes for v2

This version produces images extremely close to the anime art style. Most importantly, it rarely distorts her eyes. The key to this might be using a high repeat number and low learning rate, as well as a large enough LoRA capacity (I used 32/16). See changelog (About this version section on the left pane) for more details regarding the training parameters.

Model and ratio:

• Meina_Hentai*: around 0.6-0.7. I recommend using this since it is what the LoRA has been fine-tuned from. Do not get fooled by the name, it is actually a good SFW model

• Anything v4.5, AOMv3: similar to the above. Also precisely describes Kanon, though generated images have high saturation

Demo images did not went through any highres upscaler.

Notes for v1

Model and ratio:

• Anything v4.5: use a LoRA ratio of around 0.6-0.8

• Corneos7thHeavenMix: ratio about 0.7-0.8. Good for generating anime style art

• AOM3A1: around 0.6. Might confuse left and right bangs if generate sideway images

AOM3A1 also has the tendency to change Kanon's hairstyle to hair bun, hair tie, or add random hair bow, ornament, etc. Consider adding those to the negative prompt if you don't want them to appear in your images.

Comparing v2 and v1

Left: v1, right: v2. All images are generated with Meina_Hentai and AOMv3 VAE

描述:

Update v2

This new version is trained on the same dataset as before, with 151 anime screencaps.

The difference is that it is fine-tuned from my custom checkpoint with 0.7*Meina_Hentai + 0.3*AnyLoRA. Inference should work on any model, but Meina_Hentai might give the best result (it is a good SFW model, don't get fooled by the name).

The model is trained with below parameters and repeat 8:

[additional_network_arguments]

unet_lr = 0.00025

text_encoder_lr = 0.0001

network_dim = 32

network_alpha = 16

network_module = "networks.lora"



[optimizer_arguments]

learning_rate = 0.00025

lr_scheduler = "cosine_with_restarts"

lr_scheduler_num_cycles = 3

lr_warmup_steps = 422

optimizer_type = "AdamW8bit"



[training_arguments]

max_train_epochs = 14

save_every_n_epochs = 2

save_last_n_epochs = 10

train_batch_size = 2

clip_skip = 2

min_snr_gamma = 5.0

weighted_captions = false

seed = 42

max_token_length = 225

xformers = true

lowram = true

max_data_loader_n_workers = 8

persistent_data_loader_workers = true

save_precision = "fp16"

mixed_precision = "fp16"

output_dir = "/content/drive/MyDrive/lora_training/output/shibukano"

logging_dir = "/content/drive/MyDrive/lora_training/log"

output_name = "shibukano"

log_prefix = "shibukano"

save_state = false



[model_arguments]

pretrained_model_name_or_path = "/content/ashleymix-v2.safetensors"

v2 = false



[saving_arguments]

save_model_as = "safetensors"



[dreambooth_arguments]

prior_loss_weight = 1.0



[dataset_arguments]

cache_latents = true

The final checkpoint is created by averaging the checkpoints of epoch 12 and 8: 0.8*epoch_12 + 0.2*epoch_8.

训练词语: shibuya kanon