オルフェーヴルのLoRAはすぐに誰か作るだろうなのでオルフェーヴルではないこれから思い出されることも少なくなるだろうマスクド三冠馬を作ったKohaku-XL-Deltaベースでやったら割とうまくいった"Someone will probably create Orfevre's Lora soon. Therefore, it will not be remembered much as Orfevre in the future, having created a Triple Crown winner while masked. It worked rather well when done based on the Kohaku-XL-Delta base."prompt<pre><code>1girl,masked_sankanba,(umamusume),horse tail,horse ears, surgical mask</code></pre>v1.0意外とうまく行ったので公開版細かいディテールは潰れがち"It went surprisingly well, so here's the public version. The finer details tend to get crushed."v2.0画像枚数を増やしてgradient_checkpointingの使用､オプティマイザの変更などKohakuDELTAのトレーニングパラメーターを参考に変更潰れがちだった細かいディテールの表現力アップ"Increased the number of images and made changes such as using gradient checkpointing and changing the optimizer, referencing the training parameters of KohakuDELTA. This improved the expressiveness of the finer details that tended to get crushed."v2.0PonyAutismMix Ponyでのテスト

masked sankanba-XL マスクド三冠馬 (UmaMusume??)

<pre><code>accelerate launch 
--num_cpu_threads_per_process=2 "./sdxl_train_network.py" 
--bucket_reso_steps=64 
--cache_latents 
--cache_latents_to_disk 
--caption_dropout_rate="0.05" 
--caption_extension=".txt" 
--enable_bucket 
--min_bucket_reso=256 
--max_bucket_reso=2048 
--fp8_base 
--gradient_checkpointing 
--keep_tokens="2" 
--learning_rate="0.0003" 
--logging_dir="H:/lora/masked_sankanba-XL/log" 
--lr_scheduler="constant" 
--lr_scheduler_num_cycles="50" 
--max_data_loader_n_workers="0" 
--max_grad_norm="1" 
--resolution="1024,1024" 
--max_train_epochs=50
--max_train_steps="3834" 
--min_snr_gamma=5 
--mixed_precision="bf16" 
--network_alpha="32" 
--network_dim=32 
--network_module=networks.lora 
--no_half_vae 
--optimizer_type="AdamW" 
 --shuffle_caption 
--text_encoder_lr=0.0003 
--train_batch_size="3" 
--unet_lr=0.0003
--xformers</code></pre>

Files with SHA256: 9d0470fff4e6c3d1f608bd10d76fbcf9c4f97fd093016f740aed1eb338166cf7