The model is trained on the style of Kaisen Chuui's later works, focusing primarily on a specific uniform. The training dataset also includes some NSFW and synthetically generated images for variety.

Does not require a specific trigger word.

The recommended weight is 1, but it can be varied from 0.1 to 1.5.

The base model: Anima base-v1.0.

Automatic tag captions for images: SmilingWolf/wd-vit-large-tagger-v3 (pony/Illustrious/Danbooru-style tags)

The model was trained with random tag dropout: 0.1.

Mixed res training on [512, 1024]

The kohya-ss/sd-scripts repository was used for training. The first pass consists of 40 epochs with a high learning rate of 6e-4, and the second of 40 epochs with a lower rate of 2e-5 for stabilization.

Params:

accelerate launch --num_cpu_threads_per_process 1 anima_train_network.py `
  --pretrained_model_name_or_path="..\models\diffusion_models\anima_baseV10.safetensors" `
  --qwen3="..\models\text_encoders\qwen_3_06b_base.safetensors" `
  --vae="..\models\vae\qwen_image_vae.safetensors" `
  --dataset_config="..index.toml" `
  --output_dir="output" `
  --output_name="kaisen_chuui" `
  --save_model_as=safetensors `
  --network_module=networks.lora_anima `
  --network_dim=64 `
  --learning_rate=2e-5 `
  --optimizer_type="AdamW8bit" `
  --lr_scheduler="constant" `
  --timestep_sampling="sigmoid" `
  --discrete_flow_shift=1.0 `
  --max_train_epochs=45 `
  --save_every_n_epochs=1 `
  --mixed_precision="bf16" `
  --gradient_checkpointing `
  --cache_latents `
  --vae_chunk_size=64 `
  --network_train_unet_only `
  --vae_disable_cache `
  --network_dropout 0.1 `
  --sample_at_first `
  --sample_every_n_epochs 1 `
  --sample_prompts="..\promt.txt"