You can now run this cell to cook your Lora. Good luck!<pre><code>✅ Dependencies already installed.

💿 Checking dataset...
📁MyDrive/Loras/example/dataset
📈 Found 61 images with 4 repeats, equaling 244 steps.
📉 Divide 244 steps by 4 batch size to get 61.0 steps per epoch.
🔮 There will be 8000 steps, divided into 131 epochs and then some.

📄 Config saved to /content/drive/MyDrive/Loras/example/training_config.toml
📄 Dataset config saved to /content/drive/MyDrive/Loras/example/dataset_config.toml

⭐ Starting trainer...

Flax classes are deprecated and will be removed in Diffusers v1.0.0. We recommend migrating to PyTorch classes or pinning your version of Diffusers.
Flax classes are deprecated and will be removed in Diffusers v1.0.0. We recommend migrating to PyTorch classes or pinning your version of Diffusers.
Loading settings from /content/drive/MyDrive/Loras/example/training_config.toml...
Loading dataset config from /content/drive/MyDrive/Loras/example/dataset_config.toml
prepare images.
get image size from name of cache files
100% 61/61 [00:00&lt;00:00, 169326.63it/s]
set image size from cache files: 61/61
found directory /content/drive/MyDrive/Loras/example/dataset contains 61 image files
read caption: 100% 61/61 [00:00&lt;00:00, 238.41it/s]
244 train images with repeats.
0 reg images with repeats.
no regularization images / 正則化画像が見つかりませんでした
[Dataset 0]
 batch_size: 4
 resolution: (1024, 1024)
 skip_image_resolution: None
 resize_interpolation: None
 enable_bucket: True
 min_bucket_reso: 256
 max_bucket_reso: 4096
 bucket_reso_steps: 64
 bucket_no_upscale: False

 [Subset 0 of Dataset 0]
 image_dir: "/content/drive/MyDrive/Loras/example/dataset"
 image_count: 61
 num_repeats: 4
 shuffle_caption: False
 keep_tokens: 1
 caption_dropout_rate: 0.0
 caption_dropout_every_n_epochs: 0
 caption_tag_dropout_rate: 0.0
 caption_prefix: None
 caption_suffix: None
 color_aug: False
 flip_aug: False
 face_crop_aug_range: None
 random_crop: False
 token_warmup_min: 1,
 token_warmup_step: 0,
 alpha_mask: False
 resize_interpolation: None
 custom_attributes: {}
 is_reg: False
 class_tokens: None
 caption_extension: .txt


[Prepare dataset 0]
loading image sizes.
100% 61/61 [00:00&lt;00:00, 1070512.74it/s]
make buckets
number of images (including repeats) / 各bucketの画像枚数（繰り返し回数を含む）
bucket 0: resolution (576, 1664), count: 4
bucket 1: resolution (640, 1600), count: 4
bucket 2: resolution (704, 1472), count: 4
bucket 3: resolution (768, 1280), count: 4
bucket 4: resolution (768, 1344), count: 4
bucket 5: resolution (832, 1216), count: 80
bucket 6: resolution (896, 1152), count: 44
bucket 7: resolution (960, 1088), count: 8
bucket 8: resolution (1024, 1024), count: 32
bucket 9: resolution (1088, 960), count: 12
bucket 10: resolution (1152, 896), count: 8
bucket 11: resolution (1216, 832), count: 16
bucket 12: resolution (1280, 768), count: 12
bucket 13: resolution (1344, 768), count: 12
mean ar error (without repeats): 0.02945754669738088
preparing accelerator
accelerator device: cuda
Loading Qwen3 text encoder...
Loading Qwen3 text encoder from /content/kohya-trainer/model/split_files/text_encoders/qwen_3_06b_base.safetensors
Loaded Qwen3 state dict: &lt;All keys matched successfully&gt;
Loaded Qwen3 text encoder. Parameters: 596,049,920
Loading Anima VAE...
Initializing VAE
Loading VAE from /content/kohya-trainer/model/split_files/vae/qwen_image_vae.safetensors
Converted ComfyUI AutoencoderKL state dict keys to official format
Loaded VAE: &lt;All keys matched successfully&gt;
[Dataset 0]
caching latents with caching strategy.
caching latents...
100% 61/61 [00:00&lt;00:00, 451.81it/s]
move text encoder to gpu
[Dataset 0]
caching Text Encoder outputs with caching strategy.
checking cache validity...
100% 61/61 [00:00&lt;00:00, 838860.80it/s]
caching Text Encoder outputs...
100% 16/16 [00:03&lt;00:00, 4.58it/s]
move text encoder back to cpu
Loading Anima DiT model with attn_mode=torch, split_attn: False...
Loading DiT model from /content/kohya-trainer/model/split_files/diffusion_models/anima-base-v1.0.safetensors, device=cuda
Loading model files: ['/content/kohya-trainer/model/split_files/diffusion_models/anima-base-v1.0.safetensors']
Loading state dict without FP8 optimization. Dtype of weight: torch.float16, hook enabled: False
Loaded DiT model from /content/kohya-trainer/model/split_files/diffusion_models/anima-base-v1.0.safetensors, unexpected missing keys: 0, unexpected keys: 0
import network module: networks.lora_anima
create LoRA network. base dim (rank): 32, alpha: 16
neuron dropout: p=None, rank dropout: p=None, module dropout: p=None
create LoRA for Text Encoder 1:
create LoRA for Text Encoder 1: 196 modules.
create LoRA for Anima DiT: 280 modules.
enable LoRA for DiT: 280 modules
prepare optimizer, data loader etc.
use 8-bit AdamW optimizer | {'weight_decay': 0.1, 'betas': [0.9, 0.99]}
enable full fp16 training.
running training / 学習開始
 num train images * repeats / 学習画像の数×繰り返し回数: 244
 num validation images * repeats / 学習画像の数×繰り返し回数: 0
 num reg images / 正則化画像の数: 0
 num batches per epoch / 1epochのバッチ数: 61
 num epochs / epoch数: 132
 batch size per device / バッチサイズ: 4
 gradient accumulation steps / 勾配を合計するステップ数 = 1
 total optimization steps / 学習ステップ数: 8000
text_encoder is not needed for training. deleting to save memory.
unet dtype: torch.float16, device: cuda:0
steps: 0% 0/8000 [00:00&lt;?, ?it/s]
epoch 1/132

epoch is incremented. current_epoch: 0, epoch: 1
epoch is incremented. current_epoch: 0, epoch: 1
steps: 1% 61/8000 [25:16&lt;54:49:39, 24.86s/it, avr_loss=0.0707]
saving checkpoint: /content/drive/MyDrive/Loras/example/output/cutestyle-01.safetensors

epoch 2/132

epoch is incremented. current_epoch: 1, epoch: 2
epoch is incremented. current_epoch: 1, epoch: 2
steps: 2% 122/8000 [50:41&lt;54:32:49, 24.93s/it, avr_loss=0.0659]
saving checkpoint: /content/drive/MyDrive/Loras/example/output/cutestyle-02.safetensors

epoch 3/132

epoch is incremented. current_epoch: 2, epoch: 3
epoch is incremented. current_epoch: 2, epoch: 3
steps: 2% 183/8000 [1:16:07&lt;54:12:02, 24.96s/it, avr_loss=0.0668]
saving checkpoint: /content/drive/MyDrive/Loras/example/output/cutestyle-03.safetensors

epoch 4/132

epoch is incremented. current_epoch: 3, epoch: 4
epoch is incremented. current_epoch: 3, epoch: 4
steps: 3% 244/8000 [1:41:36&lt;53:49:42, 24.98s/it, avr_loss=0.064]
saving checkpoint: /content/drive/MyDrive/Loras/example/output/cutestyle-04.safetensors

epoch 5/132

epoch is incremented. current_epoch: 4, epoch: 5
epoch is incremented. current_epoch: 4, epoch: 5
steps: 4% 305/8000 [2:07:04&lt;53:25:59, 25.00s/it, avr_loss=0.0643]
saving checkpoint: /content/drive/MyDrive/Loras/example/output/cutestyle-05.safetensors

epoch 6/132

epoch is incremented. current_epoch: 5, epoch: 6
epoch is incremented. current_epoch: 5, epoch: 6
steps: 5% 366/8000 [2:32:34&lt;53:02:27, 25.01s/it, avr_loss=0.0609]
saving checkpoint: /content/drive/MyDrive/Loras/example/output/cutestyle-06.safetensors

epoch 7/132

epoch is incremented. current_epoch: 6, epoch: 7
epoch is incremented. current_epoch: 6, epoch: 7
steps: 5% 427/8000 [2:58:05&lt;52:38:32, 25.02s/it, avr_loss=0.0676]
saving checkpoint: /content/drive/MyDrive/Loras/example/output/cutestyle-07.safetensors

epoch 8/132

epoch is incremented. current_epoch: 7, epoch: 8
epoch is incremented. current_epoch: 7, epoch: 8
steps: 6% 488/8000 [3:23:37&lt;52:14:32, 25.04s/it, avr_loss=0.063]
saving checkpoint: /content/drive/MyDrive/Loras/example/output/cutestyle-08.safetensors

epoch 9/132

epoch is incremented. current_epoch: 8, epoch: 9
epoch is incremented. current_epoch: 8, epoch: 9
steps: 7% 531/8000 [3:41:37&lt;51:57:20, 25.04s/it, avr_loss=0.0613]</code></pre><hr />

Files with SHA256: c31d575c78eae1c30d51e29747d3f0ec304e85280f4dcdf725238f4598ce4745