I migrated this project to Fwaun Tune.

Fwaun Tune に移行しました。

A fine-tuned model specializing in flat color, no lineart, slightly chibi character art.

フラットカラー、主線なし、デフォルメ（ちびというほどじゃない）の画風を学習させたファインチューンモデルです。

How this model was made / このモデルの作り方 — fwaun-tools

This model was built with an open-source pipeline I develop, fwaun-tools. Sharing the workflow for transparency:

1. Dataset prep — tags and captions are curated with fwaun-tools (manual editing + WD14 tagger + captioning with OpenAI API + Danbooru tags in one surface).

2. Training — the base model is trained with sd-scripts on that dataset.

3. Diff extraction & merge — fwaun-tools extracts the full fine-tune delta (tuned − base) and transfers it onto a Turbo checkpoint with merge-diff.

4. Quantization — the merged checkpoint is quantized to INT8 (+ConvRot) with fwaun-tools.

fwaun-tools is a personal project — not heavily tested, so bugs are likely — but issues and pull requests are welcome.

このモデルは、自作のオープンソースツール fwaun-tools を使ったパイプラインで制作しています。透明性のためワークフローを公開します:

1. データセット準備 — タグとキャプションを fwaun-tools で整備(手動編集＋WD14タガー＋Open AI API 互換キャプショナー＋Danbooruタグを1画面に統合)。

2. 学習 — そのデータセットで sd-scripts を使ってベースモデルを学習。

3. 差分抽出・マージ — fwaun-tools の merge-diff でフルファインチューンの差分(tuned − base)を抽出し、Turbo チェックポイントへ転写。

4. 量子化 — マージ後のチェックポイントを fwaun-tools で INT8(+ConvRot)に量子化。

fwaun-tools は個人開発のツールです。テストは十分ではなく不具合の可能性がありますが、イシューやプルリクエストは歓迎します。

v4.0

• Change trigger words.

• Update dataset to increase the character's head-to-body ratio

v3.0

To help body proportions, I’ve added super deformed, chibi-style illustrations to the training data.

You can decrease the body-to-head ratio using the “chibi, super deformed” tags and you can also use "realistic" tag to increase it.

Trigger word

flat color, no lineart, deformed

Optional tags

chibi, super deformed

realistic

v2.0

To help adjust color tones and body proportions, I’ve added non-deformed flat color no lineart illustrations to the training data.

You can increase the body-to-head ratio using the “realistic” tag. If you don't want to increase it, set it to a negative value.

Trigger word

flat color, no lineart, deformed

Negative word

chibi, realistic

Initial release

Note: This is my first fine-tuned model. While I'm pleased with this results, the quality may vary compared to established style LoRAs. I appreciate your patience and feedback as I continue to refine this approach. Your input helps me improve future versions!

This model excels at creating characters with a distinctive light deformation style - more expressive than realistic, yet more refined than super-deformed (SD/chibi). Perfect for characters ranging from 3.5 to 7 heads tall, capturing that sweet spot between youthful charm and mature appeal.

Key characteristics:

Clean flat colors (no lineart required)

Soft, natural deformation - ideal for stylized character design

Works beautifully with other character LoRAs to apply different character traits while keeping this style intact

Stable on deformed proportions that danbooru tags struggle to capture

Recommended usage: Combine with character LoRAs for maximum flexibility. The model's style strength is intentionally calibrated to enhance, not overwhelm, other adapters.

Training Details

For transparency and to help with troubleshooting, here's the training command I used with sd-scripts:

accelerate launch --num_cpu_threads_per_process 1 anima_train.py --pretrained_model_name_or_path=/path/to/animaOfficial_preview3Base.safetensors \
	--qwen3=/path/to/qwen_3_06b_base.safetensors \
	--vae=/path/to/qwen_image_vae.safetensors \
	--dataset_config=/path/to/dataset_config_anima.toml \
	--output_dir=/path/to/outputu \
	--output_name=fwaunstp_ckpt_anima_preview3_v1 \
	--save_model_as=safetensors \
	--learning_rate=1e-6 \
	--optimizer_type=AdamW8bit \
	--lr_scheduler=constant \
	--timestep_sampling=sigmoid \
	--discrete_flow_shift=1.0 \
	--max_train_epochs=8 \
	--save_every_n_epochs=1 \
	--mixed_precision=bf16 \
	--gradient_checkpointing \
	--cache_latents \
	--vae_chunk_size=64 \
	--vae_disable_cache \
	--save_state \
	--split_attn \
	--attn_mode=torch

This model was trained using f32 precision (not BF16, which some other Anima Preview 3 models use). If you notice any issues or compatibility problems, please let me know—I'd appreciate feedback on the training parameters or the checkpoint format.