前回はKohya ssでLoRAを自作したものの、出来上がったLoRAの質は低いものだったね。
元の姿をほとんど学習していない様子だった。
今回は学習ツールを変更してAnima LoRA Factoryを利用してみようと思ったよ。
▼参考にした記事
https://note.com/cocoat/n/n63633d631148
画面構成や項目はKohya ssに比べて見やすい。
画像素材を指定し、
タグを設定してーー
checkpointやvaeやtext encoderを設定して、出力先と生成loRA名を選ぶ。
あとは、ちょっとした設定をしたら学習開始。
学習プロセスを開始しています...
--- Connection Established ---
[TRAIN] 2026-09-16 16:29:47 INFO Loading dataset config from train_network.py:502
[TRAIN] C:\Users\xxx\Anima-LoRA-Factory-v
[TRAIN] 5.1\backend\dataset_config.to
[TRAIN] ml
[TRAIN] INFO prepare images. train_util.py:2141
[TRAIN] INFO get image size from name of train_util.py:2016
[TRAIN] cache files
[TRAIN]
[TRAIN] 0%| | 0/33 [00:00<?, ?it/s]
[TRAIN] 100%|██████████| 33/33 [00:00<?, ?it/s]
[TRAIN] INFO set image size from cache train_util.py:2046
[TRAIN] files: 0/33
[TRAIN] INFO found directory train_util.py:2088
[TRAIN] C:\Users\xxx\im
[TRAIN] g_src\10_fum2asahi contains 33
[TRAIN] image files
[TRAIN]
[TRAIN] read caption: 0%| | 0/33 [00:00<?, ?it/s]
[TRAIN] read caption: 100%|██████████| 33/33 [00:00<00:00, 5251.23it/s]
[TRAIN] INFO 132 train images with repeats. train_util.py:2187
[TRAIN] INFO 0 reg images with repeats. train_util.py:2191
[TRAIN] WARNING no regularization images / train_util.py:2196
[TRAIN] 正則化画像が見つかりませんでし
[TRAIN] た
[TRAIN] INFO [Dataset 0] config_util.py:584
[TRAIN] batch_size: 1
[TRAIN] resolution: (1024, 1024)
[TRAIN] skip_image_resolution: None
[TRAIN] resize_interpolation: None
[TRAIN] enable_bucket: True
[TRAIN] min_bucket_reso: 256
[TRAIN] max_bucket_reso: 2048
[TRAIN] bucket_reso_steps: 64
[TRAIN] bucket_no_upscale: False
[TRAIN]
[TRAIN] [Subset 0 of Dataset 0]
[TRAIN] image_dir:
[TRAIN] "C:\Users\xxx\im
[TRAIN] g_src\10_fum2asahi"
[TRAIN] image_count: 33
[TRAIN] num_repeats: 4
[TRAIN] shuffle_caption: False
[TRAIN] keep_tokens: 0
[TRAIN] caption_dropout_rate: 0.0
[TRAIN] caption_dropout_every_n_epo
[TRAIN] chs: 0
[TRAIN] caption_tag_dropout_rate:
[TRAIN] 0.0
[TRAIN] caption_prefix: None
[TRAIN] caption_suffix: None
[TRAIN] color_aug: False
[TRAIN] flip_aug: False
[TRAIN] face_crop_aug_range: None
[TRAIN] random_crop: False
[TRAIN] token_warmup_min: 1,
[TRAIN] token_warmup_step: 0,
[TRAIN] alpha_mask: False
[TRAIN] resize_interpolation: None
[TRAIN] custom_attributes: {}
[TRAIN] is_reg: False
[TRAIN] class_tokens: None
[TRAIN] caption_extension: .txt
[TRAIN]
[TRAIN]
[TRAIN] INFO [Prepare dataset 0] config_util.py:596
[TRAIN] INFO loading image sizes. train_util.py:995
[TRAIN]
[TRAIN] 0%| | 0/33 [00:00<?, ?it/s]
[TRAIN] 100%|██████████| 33/33 [00:00<00:00, 8885.67it/s]
[TRAIN] INFO make buckets train_util.py:1018
[TRAIN] INFO number of images (including train_util.py:1064
[TRAIN] repeats) /
[TRAIN] 各bucketの画像枚数（繰り返し回
[TRAIN] 数を含む）
[TRAIN] INFO bucket 0: resolution (1344, train_util.py:1069
[TRAIN] 768), count: 132
[TRAIN] INFO mean ar error (without train_util.py:1077
[TRAIN] repeats): 0.028666666666666618
[TRAIN] INFO preparing accelerator train_network.py:576
[TRAIN] INFO Loading Qwen3 text anima_train_network.py:86
[TRAIN] encoder...
[TRAIN] INFO Loading Qwen3 text encoder from anima_utils.py:201
[TRAIN] C:/Users/xxx/qwen_3_0
[TRAIN] 6b_base.safetensors
[TRAIN] 2026-09-16 16:29:51 INFO Loaded Qwen3 state dict: <All anima_utils.py:248
[TRAIN] keys matched successfully>
[TRAIN] INFO Loaded Qwen3 text encoder. anima_utils.py:256
[TRAIN] Parameters: 596,049,920
[TRAIN] 2026-09-16 16:29:52 INFO Loading Anima VAE... anima_train_network.py:91
[TRAIN] INFO Initializing VAE qwen_image_autoencoder_kl.py:1609
[TRAIN] INFO Loading VAE from qwen_image_autoencoder_kl.py:1631
[TRAIN] C:/Users/xxx/anima
[TRAIN] _qwen_image_vae.
[TRAIN] safetensors
[TRAIN] INFO Converted qwen_image_autoencoder_kl.py:1538
[TRAIN] ComfyUI
[TRAIN] AutoencoderKL
[TRAIN] state dict keys
[TRAIN] to official
[TRAIN] format
[TRAIN] INFO Loaded VAE: <All qwen_image_autoencoder_kl.py:1638
[TRAIN] keys matched
[TRAIN] successfully>
[TRAIN] INFO Loading Anima DiT model anima_train_network.py:112
[TRAIN] with attn_mode=torch,
[TRAIN] split_attn: False...
[TRAIN] 2026-09-16 16:29:53 INFO Loading DiT model from anima_utils.py:105
[TRAIN] C:/Users/xxx/anim
[TRAIN] a_baseV10.safetensors,
[TRAIN] device=cpu
[TRAIN] INFO Loading model files: lora_utils.py:90
[TRAIN] ['C:/Users/xxx/anima_ba
[TRAIN] seV10.safetensors']
[TRAIN] INFO Loading state dict without FP8 lora_utils.py:264
[TRAIN] optimization. Dtype of weight:
[TRAIN] torch.bfloat16, hook enabled:
[TRAIN] False
[TRAIN] accelerator device: cuda
[TRAIN]
[TRAIN] Loading anima_baseV10.safetensors: 0%| | 0/685 [00:00<?, ?it/s]
[TRAIN] Loading anima_baseV10.safetensors: 5%|▍ | 33/685 [00:00<00:02, 323.96it/s]
[TRAIN] Loading anima_baseV10.safetensors: 10%|█ | 71/685 [00:00<00:01, 352.95it/s]
[TRAIN] Loading anima_baseV10.safetensors: 16%|█▌ | 107/685 [00:00<00:01, 340.87it/s]
[TRAIN] Loading anima_baseV10.safetensors: 21%|██ | 142/685 [00:00<00:01, 334.46it/s]
[TRAIN] Loading anima_baseV10.safetensors: 26%|██▌ | 176/685 [00:00<00:01, 329.71it/s]
[TRAIN] Loading anima_baseV10.safetensors: 31%|███ | 213/685 [00:00<00:01, 342.77it/s]
[TRAIN] Loading anima_baseV10.safetensors: 37%|███▋ | 253/685 [00:00<00:01, 343.95it/s]
[TRAIN] Loading anima_baseV10.safetensors: 43%|████▎ | 293/685 [00:00<00:01, 346.99it/s]
[TRAIN] Loading anima_baseV10.safetensors: 49%|████▊ | 333/685 [00:00<00:01, 347.01it/s]
[TRAIN] Loading anima_baseV10.safetensors: 54%|█████▍ | 373/685 [00:01<00:00, 350.96it/s]
[TRAIN] Loading anima_baseV10.safetensors: 60%|██████ | 413/685 [00:01<00:00, 348.41it/s]
[TRAIN] Loading anima_baseV10.safetensors: 66%|██████▌ | 453/685 [00:01<00:00, 356.73it/s]
[TRAIN] Loading anima_baseV10.safetensors: 72%|███████▏ | 492/685 [00:01<00:00, 356.57it/s]
[TRAIN] Loading anima_baseV10.safetensors: 77%|███████▋ | 528/685 [00:01<00:00, 352.81it/s]
[TRAIN] Loading anima_baseV10.safetensors: 82%|████████▏ | 564/685 [00:01<00:00, 349.07it/s]
[TRAIN] Loading anima_baseV10.safetensors: 98%|█████████▊| 668/685 [00:01<00:00, 542.49it/s]
[TRAIN]
[TRAIN] 2026-09-16 16:29:54 INFO Loaded DiT model from anima_utils.py:146
[TRAIN] C:/Users/xxx/anim
[TRAIN] a_baseV10.safetensors,
[TRAIN] unexpected missing keys: 0,
[TRAIN] unexpected keys: 0
[TRAIN] INFO enable block swap: anima_train_network.py:131
[TRAIN] blocks_to_swap=20
[TRAIN] INFO Anima: Block swap enabled. anima_models.py:1262
[TRAIN] Swapping 20 blocks, total
[TRAIN] blocks: 28, device: cuda.
[TRAIN] INFO create LoRA network. base dim lora_anima.py:426
[TRAIN] (rank): 16, alpha: 16.0
[TRAIN] INFO neuron dropout: p=None, rank lora_anima.py:427
[TRAIN] dropout: p=None, module dropout:
[TRAIN] p=None
[TRAIN] INFO create LoRA for Text Encoder 1: lora_anima.py:531
[TRAIN] INFO create LoRA for Text Encoder 1: lora_anima.py:533
[TRAIN] 196 modules.
[TRAIN] 2026-09-16 16:29:55 INFO create LoRA for Anima DiT: 280 lora_anima.py:545
[TRAIN] modules.
[TRAIN] INFO enable LoRA for text encoder: lora_anima.py:584
[TRAIN] 196 modules
[TRAIN] INFO enable LoRA for DiT: 280 modules lora_anima.py:589
[TRAIN] INFO Text Encoder 1 (Qwen3): 196 lora_anima.py:717
[TRAIN] modules, LR 5e-05
[TRAIN] INFO use AdamW optimizer | {} train_util.py:5159
[TRAIN] fatal: not a git repository (or any of the parent directories): .git
[TRAIN] 2026-09-16 16:29:59 INFO unet dtype: torch.bfloat16, train_network.py:1332
[TRAIN] device: cuda:0
[TRAIN] INFO text_encoder [0] dtype: train_network.py:1338
[TRAIN] torch.bfloat16, device:
[TRAIN] cuda:0
[TRAIN] import network module: networks.lora_anima
[TRAIN] prepare optimizer, data loader etc.
[TRAIN] override steps. steps for 10 epochs is / 指定エポックまでのステップ数: 1320
[TRAIN] running training / 学習開始
[TRAIN] num train images * repeats / 学習画像の数×繰り返し回数: 132
[TRAIN] num validation images * repeats / 学習画像の数×繰り返し回数: 0
[TRAIN] num reg images / 正則化画像の数: 0
[TRAIN] num batches per epoch / 1epochのバッチ数: 132
[TRAIN] num epochs / epoch数: 10
[TRAIN] batch size per device / バッチサイズ: 1
[TRAIN] gradient accumulation steps / 勾配を合計するステップ数 = 1
[TRAIN] total optimization steps / 学習ステップ数: 1320
[TRAIN]
[TRAIN] steps: 0%| | 0/1320 [00:00<?, ?it/s] INFO epoch is incremented. train_util.py:787
[TRAIN] current_epoch: 0, epoch: 1
[TRAIN]
[TRAIN] steps: 0%| | 1/1320 [00:05<2:08:41, 5.85s/it]
[TRAIN] steps: 0%| | 1/1320 [00:05<2:08:48, 5.86s/it, avr_loss=0.011]
[TRAIN] steps: 0%| | 2/1320 [00:10<1:59:46, 5.45s/it, avr_loss=0.011]
[TRAIN] steps: 0%| | 2/1320 [00:10<1:59:49, 5.45s/it, avr_loss=0.0115]
略
[TRAIN] steps: 100%|█████████▉| 1319/1320 [1:51:22<00:05, 5.07s/it, avr_loss=0.0102]
[TRAIN] steps: 100%|█████████▉| 1319/1320 [1:51:22<00:05, 5.07s/it, avr_loss=0.0102]
[TRAIN] steps: 100%|██████████| 1320/1320 [1:51:27<00:00, 5.07s/it, avr_loss=0.0102]
[TRAIN] steps: 100%|██████████| 1320/1320 [1:51:27<00:00, 5.07s/it, avr_loss=0.0102]2026-09-16 18:21:27 WARNING Failed to get git commit sai_model_spec.py:288
[TRAIN] hash, using fallback
[TRAIN] INFO model saved. train_network.py:1714
[TRAIN]
[TRAIN] steps: 100%|██████████| 1320/1320 [1:51:27<00:00, 5.07s/it, avr_loss=0.0102]
[TRAIN]
[TRAIN] epoch 1/10
[TRAIN]
[TRAIN]
[TRAIN] epoch 2/10
[TRAIN]
[TRAIN]
[TRAIN] epoch 3/10
[TRAIN]
[TRAIN]
[TRAIN] epoch 4/10
[TRAIN]
[TRAIN]
[TRAIN] epoch 5/10
[TRAIN]
[TRAIN]
[TRAIN] epoch 6/10
[TRAIN]
[TRAIN]
[TRAIN] epoch 7/10
[TRAIN]
[TRAIN]
[TRAIN] epoch 8/10
[TRAIN]
[TRAIN]
[TRAIN] epoch 9/10
[TRAIN]
[TRAIN]
[TRAIN] epoch 10/10
[TRAIN]
[TRAIN]
[TRAIN] saving checkpoint: C:/Users/xxx/Anima-LoRA-Factory-v5.1/output\fum2asahi2609161630_unet.safetensors
[TRAIN]
--- TRAIN Finished ---
[TRAIN]
--- Starting Conversion to ComfyUI ---
[TRAIN] 2026-09-16 18:21:31 INFO Loading source convert_anima_lora_to_comfy.py:20
[TRAIN] file
[TRAIN] C:/Users/xxx/Anima-LoRA-
[TRAIN] Factory-v5.1/out
[TRAIN] put\fum2asahi260
[TRAIN] 9161630_unet.saf
[TRAIN] etensors
[TRAIN] INFO Converting... convert_anima_lora_to_comfy.py:27
[TRAIN] INFO Converted 1428 convert_anima_lora_to_comfy.py:124
[TRAIN] keys
[TRAIN] INFO Calculating convert_anima_lora_to_comfy.py:134
[TRAIN] hashes and
[TRAIN] creating
[TRAIN] metadata...
[TRAIN] 2026-09-16 18:21:32 INFO Saving convert_anima_lora_to_comfy.py:140
[TRAIN] destination
[TRAIN] file
[TRAIN] C:/Users/xxx/Anima-Lo
[TRAIN] RA-Factory-v5.1
[TRAIN] /output\fum2asa
[TRAIN] hi2609161630_co
[TRAIN] mfy.safetensors
[TRAIN]
--- TRAIN Finished ---
[TRAIN] Cleaning up temporary UNET file: fum2asahi2609161630_unet.safetensors
[TRAIN]
--- All Processes Finished Successfully ---
出力先フォルダにLoRAが格納された！
じゃあStableDiffusionで生成してみよう。
あれ、エラー？
学習時と同じcheckpointとvaeを選んだのに？
仕方ないから違うAnima系vaeで生成。
ええ……真っ黒なの？
これじゃあKohya ssより酷い結果じゃないか……。