[SD 1.5][LORA]The Idolmaster Cinderella Girls LoRA Lite (style+9 chars from U149) v1-linear
<p>A LoRA (Hadamard Product) based on Cinderella Girls Cards and Starlight Stage Cards of the <a target="_blank" rel="ugc" href="https://cinderella-u149-anime.idolmaster-official.jp/character">9 characters in the upcoming U149 anime</a>. A small test before training for many more characters in Starlight Stage.</p><p></p><h2>Usage</h2><p>This Hadamard Product LoRA requires installing <a target="_blank" rel="ugc" href="https://github.com/KohakuBlueleaf/a1111-sd-webui-locon">an extra extension</a> to be used in the Web UI.<br /></p><p>The style itself doesn't have a trigger word. For the characters, below are the character tags and respective top 20 related tags.</p><pre><code>古賀小春,short hair, brown hair, 1girl, smile, brown eyes, solo, open mouth, bow, dress, looking at viewer, hair bow, hairband, blush, sitting, frills, light brown hair, sky, :d, gloves, blonde hair
赤城みりあ,
short hair, black hair, smile, brown eyes, two side up, open mouth, looking at viewer, 1girl, solo, blush, hair ornament, :d, brown hair, multiple girls, skirt, jewelry, one eye closed, closed eyes, gloves, shorts
結城晴,
orange hair, smile, 1girl, solo, long hair, open mouth, shorts, looking at viewer, blush, sky, day, brown eyes, gloves, outdoors, multiple girls, brown hair, midriff, navel, hat, shirt
龍崎薫,
short hair, hair ornament, hairclip, smile, brown hair, open mouth, 1girl, yellow eyes, solo, orange hair, shorts, looking at viewer, closed eyes, hat, brown eyes, blush, food, multiple girls, flower, skirt
的場梨沙,
black hair, long hair, yellow eyes, twintails, smile, jewelry, 1girl, ribbon, solo, hair ribbon, open mouth, necklace, looking at viewer, blush, skirt, animal print, bangs, midriff, one eye closed, sitting
橘ありす,
brown eyes, brown hair, long hair, 1girl, solo, bow, looking at viewer, smile, hair bow, dress, open mouth, blush, ribbon, food, flower, sitting, fruit, jewelry, hair ornament, strawberry
佐々木千枝,
short hair, black hair, smile, black eyes, hair ornament, 1girl, rabbit hair ornament, solo, open mouth, hairclip, looking at viewer, blush, skirt, dress, brown hair, jewelry, holding, multiple girls, sitting, indoors
櫻井桃華,
blonde hair, green eyes, short hair, smile, 1girl, looking at viewer, solo, blush, wavy hair, dress, open mouth, frills, hairband, flower, hair ornament, bow, gloves, jewelry, rose, sitting
市原仁奈,
long hair, brown hair, smile, open mouth, blush, 1girl, brown eyes, solo, blunt bangs, cosplay, animal costume, kigurumi, :d, bow, bangs, looking at viewer, closed eyes, gloves, outdoors, hat</code></pre><h2>Training info</h2><p><br />Trained solely on 164 640x800 frameless Cinderella Girls Cards and 273 1280x824 Starlight Stage Cards collected from <a target="_blank" rel="ugc" href="https://imas.gamedbs.jp/cg/">アイマス シンデレラガールズ カードギャラリー</a> without balancing and tag cleaning. Near-duplicates were found (minor differences like closed/open eyes), but they were not cleaned. As the jpgs are highly compressed, they are passed through <a target="_blank" rel="ugc" href="https://github.com/xinntao/Real-ESRGAN/blob/master/docs/anime_model.md">RealESRGAN_x4plus_anime_6B</a> with --outscale=1 in order to remove the artifacts. Tags generated with <a target="_blank" rel="ugc" href="https://huggingface.co/SmilingWolf/wd-v1-4-swinv2-tagger-v2">SmilingWolf/wd-v1-4-swinv2-tagger-v2</a> with a threshold of 0.35. Caption template was ", ".</p><p></p><p>Originally trained with LoRA on convolution blocks ("LoCon"), but it was found that removing the convolution weights from the output doesn't alter the results significantly. The original is available as <code>v1-full</code></p><p>and the conv-pruned version is <code>v1-linear</code></p><p></p><p>Training cost: ~4.5 T4-hours</p><p></p><p>Data distribution:</p><pre><code>Character Total 640 1280
---------------------------------
櫻井桃華 60 36 24
市原仁奈 56 38 18
橘ありす 54 34 20
赤城みりあ 53 37 16
佐々木千枝 52 30 22
龍崎薫 50 34 16
結城晴 44 26 18
的場梨沙 40 22 18
古賀小春 28 16 12</code></pre><p></p><p>LoRA meta dump:</p><pre><code>{'ss_cache_latents': 'False',
'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_dataset_dirs': '{"meta_lat.json": {"n_repeats": 1, "img_count": 437}}',
'ss_datasets': '[{"is_dreambooth": false, "batch_size_per_device": 3, "num_train_images": 437, "num_reg_images": 0, "resolution": , "enable_bucket": true, "min_bucket_reso": null, "max_bucket_reso": null, "tag_frequency": {"meta_lat.json": {}}, "bucket_info": {"buckets": {"0": {"resolution": , "count": 273}, "1": {"resolution": , "count": 164}}, "mean_img_ar_error": 0.0}, "subsets": [{"img_count": 437, "num_repeats": 1, "color_aug": false, "flip_aug": false, "random_crop": false, "shuffle_caption": true, "keep_tokens": 0, "image_dir": "train_data", "metadata_file": "meta_lat.json"}]}]',
'ss_epoch': '45',
'ss_face_crop_aug_range': 'None',
'ss_full_fp16': 'False',
'ss_gradient_accumulation_steps': '1',
'ss_gradient_checkpointing': 'False',
'ss_learning_rate': '0.0001',
'ss_lowram': 'True',
'ss_lr_scheduler': 'polynomial',
'ss_lr_warmup_steps': '100',
'ss_max_grad_norm': '1.0',
'ss_max_token_length': '225',
'ss_max_train_steps': '7300',
'ss_mixed_precision': 'fp16',
'ss_network_alpha': '16',
'ss_network_dim': '16',
'ss_network_module': 'lycoris.kohya',
'ss_new_sd_model_hash': '89d59c3dde4c56c6d5c41da34cc55ce479d93b4007046980934b14db71bdb2a8',
'ss_noise_offset': 'None',
'ss_num_batches_per_epoch': '146',
'ss_num_epochs': '50',
'ss_num_reg_images': '0',
'ss_num_train_images': '437',
'ss_optimizer': 'bitsandbytes.optim.adamw.AdamW8bit',
'ss_output_name': 'u149',
'ss_prior_loss_weight': '1.0',
'ss_sd_model_hash': '925997e9',
'ss_sd_model_name': 'Animefull-final-pruned.ckpt',
'ss_sd_scripts_commit_hash': '5b506b14a3f0cda56c0012bb502590e785f5ff21',
'ss_seed': 'None',
'ss_session_id': '505064346',
'ss_tag_frequency': '{"meta_lat.json": {}}',
'ss_text_encoder_lr': '5e-05',
'ss_training_comment': 'None',
'ss_training_finished_at': '1678689998.9046714',
'ss_training_started_at': '1678673431.1014886',
'ss_unet_lr': '0.0001',
'ss_v2': 'False',
'sshs_legacy_hash': '65cdfd7c',
'sshs_model_hash': '6a126d6984ae0903678f7a149e18a0b2ce34bf6e9ee4bb10311c3f1b82ea2901'}</code></pre>
页:
[1]