Update: SDXL 1.0 version <a rel="ugc" href="https://civarchive.com/models/376696/sdxl-shibuya-kanon-lora?modelVersionId=420658">here</a>, 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 <code>from side</code>). This is important, as previous models I trained for Kanon as well as other characters lacked the ability to generate sideway images.Activation tag: <code>shibuya kanon</code>To generate school uniform: <code>student uniform, gray dress, red ribbon, neck ribbon, pinafore dress, white shirt</code>To generate glasses: <code>glasses</code>. For v2, include <code>bangs, hair between eyes</code> as well (the model will generate her wearing glasses without bangs if not prompted)<h2 id="notes-for-v2">Notes for v2</h2>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:<ul><li>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</li><li>Anything v4.5, AOMv3: similar to the above. Also precisely describes Kanon, though generated images have high saturation</li></ul>Demo images did not went through any highres upscaler.<h2 id="notes-for-v1">Notes for v1</h2>Model and ratio:<ul><li>Anything v4.5: use a LoRA ratio of around 0.6-0.8</li><li>Corneos7thHeavenMix: ratio about 0.7-0.8. Good for generating anime style art</li><li>AOM3A1: around 0.6. Might confuse left and right bangs if generate sideway images</li></ul>AOM3A1 also has the tendency to change Kanon's hairstyle to <code>hair bun</code>, <code>hair tie</code>, 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.<h2 id="comparing-v2-and-v1">Comparing v2 and v1</h2>Left: v1, right: v2. All images are generated with Meina_Hentai and AOMv3 VAE<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/61c06c3b-5443-4dda-9e26-b5e908f64bab/width=525/61c06c3b-5443-4dda-9e26-b5e908f64bab.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/350cdc16-b19d-4446-888b-c26178ec51f2/width=525/350cdc16-b19d-4446-888b-c26178ec51f2.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e9e28c14-cfc2-43ee-9653-12c764e25bb9/width=525/e9e28c14-cfc2-43ee-9653-12c764e25bb9.jpeg" />

[LoRA] Shibuya Kanon (Love Live! Superstar!!)

Update v2This 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:<pre><code>[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
</code></pre>The final checkpoint is created by averaging the checkpoints of epoch 12 and 8: 0.8*epoch_12 + 0.2*epoch_8.

Files with SHA256: 8eebb572c66b48da87d27611c975b841f6b83dba1b779b7557e51a6f57e10876