【EXPERIMENTAL】Li Wen (WILL: A Wonderful World) 李雯 (WILL：美好世界)

Li Wen (WILL: A Wonderful World) 李雯 (WILL：美好世界)

Alicia (WILL: A Wonderful World) 阿莉西亚 (WILL：美好世界)

Pi (WILL: A Wonderful World) Pi (WILL：美好世界)

Triggers with liwen_will, weight has to be 0.6.

This LoRA is an experiment to answer whether I can make a decent LoRA with only 10 copies of the same image, considering that Li Wen unfortunately doesn't have that many artworks to put into the dataset, even with Lofter and TapTap's gallery.

The dataset was this:

As you can see... it's the same image. I didn't used regularization images, but I might as I try to use the AI outputs as inputs for a future v2.

Used Cetus-mix and Anypastel for example images.

Usage

General tags: liwen_will works fine already, but you can use pink hair, long hair, orange eyes, long hair if necessary.

For her uniform, just use default_outfit, school uniform, black vest, black skirt, yellow ribbon, long sleeves, white sleeves.

Known Issues

• Don't bother using it at a weight higher than 0.6, it'll overfit.

• Sometimes her sleeves will appear as black. Sometimes I tag long sleeves, white sleeves to make it happen less.

Training Parameters

• "ss_text_encoder_lr":"1e-05",

• "ss_total_batch_size":"2",

• "ss_num_epochs":"10",

• "ss_unet_lr":"0.0001",

• "ss_learning_rate":"0.0001",

• "ss_lr_scheduler":"cosine_with_restarts",

• "ss_lr_warmup_steps":"0",

• "ss_max_token_length":"225",

• "ss_mixed_precision":"fp16",

• "ss_network_alpha":"64.0",

• "ss_network_dim":"128",

• {\"n_repeats\": 20, \"img_count\": 10}

• "ss_resolution":"(768, 768)"

• "ss_sd_model_name":"Animefull-final-pruned.ckpt",

• "ss_seed":"42",