AnimeNSFW-testing

AnimeNSFW-testing

This is a LoRA for Qwen Image Edit 2511, trained solely on NSFW images with Danbooru tags collected from the internet, without using any control images.

The basic usage of Qwen Image Edit involves describing the desired edits in text with reference to a reference image, but Danbooru tags can also be used to some extent.

This LoRA was created as an experiment to test how far training can go using only Danbooru tags, as well as how effective training can be without control images, so it does not have much practical utility. It also serves as an experiment to attempt high-resolution, high-DIM training on an RTX PRO 6000, which is difficult to achieve on a 5090.

How to Use

Please incorporate AnimeNSFW.safetensors into the workflow of Qwen Image Edit 2511.

You will need the uncensored version of the text encoder, so please download it.

https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-abliterated-GGUF/tree/main

For the workflow, drag and drop the image from the following example into ComfyUI (Note: the random seed control "fixed").

Example Outputs

• 1girl, solo, standing, park, naughty, grin, looking at viewer, school uniform, blazer, jacket, shirt, skirt, tie, anime coloring (No LoRA, no control images, Danbooru tags)

• 1girl, solo, standing, park, naughty, grin, looking at viewer, school uniform, blazer, jacket, shirt, skirt, tie, anime coloring (LoRA strength 0.8, no control images, Danbooru tags)

• A young girl in a school uniform, including a blazer and skirt, is standing alone in a park. She's wearing a tie and has a naughty grin, looking directly at the viewer with anime-like coloring. (No LoRAなし, no control images, English text)

• A young girl in a school uniform, including a blazer and skirt, is standing alone in a park. She's wearing a tie and has a naughty grin, looking directly at the viewer with anime-like coloring. (LoRA strength 0.8, no control images, English text)

• 1girl, solo, standing, park, naughty, grin, looking at viewer, school uniform, blazer, jacket, shirt, skirt, tie, skirt lift by both hands, pink panties, panties, underwear, anime coloring (LoRA strength 1.0, no control images, Danbooru tags)

• A girl, dressed in a school uniform consisting of a blazer, jacket, shirt, and skirt, is standing alone in a park. She has a naughty grin and is looking directly at the viewer. The skirt is lifted by both hands, revealing pink panties underneath. The image is rendered in anime-style coloring. (LoRA strength 1.0, no control images, English text)

• control image

• 1girl, skirt lift by both hands, pussy, cleft of venus, uncensored (LoRA strength 1.0, no control images, Danbooru tags, no workflow)

(mosaic censorship manually added)

Training Method

The training method is described in Japanese below.

AnimeNSFW-testing

これは、コントロール画像を用いず、ネットで収集したdanbooruタグ付きのNSFW画像のみから学習したQwen Image Edit 2511用のLoRAです。

Qwen Image Editは参照画像に対して文章で編集内容用を記述するのが基本ですが、danbooruタグもある程度使えます。danbooruタグのみでどこまで学習できるかの実験とコントロール画像なしの場合にどこまで学習できるかの実験なので、このLoRAに実用性はあまりありません。5090では難しい高解像度・高DIMでの学習をRTX PRO 6000で試す実験も兼ねています。

使い方

Qwen Image Edit 2511のワークフローにAnimeNSFW.safetensorsを組み入れてください。 検閲解除版テキストエンコーダが必要なのでダウンロードしてください。

https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-abliterated-GGUF/tree/main

ワークフローは次の実行例の画像をComfyUIにドラッグアンドドロップしてください(乱数制御がfixedになっているので気をつけてください)。

実行例

• 1girl, solo, standing, park, naughty, grin, looking at viewer, school uniform, blazer, jacket, shirt, skirt, tie, anime coloring (LoRAなし、コントロール画像なし、 danbooruタグ)

• 1girl, solo, standing, park, naughty, grin, looking at viewer, school uniform, blazer, jacket, shirt, skirt, tie, anime coloring (LoRA強度0.8、コントロール画像なし、 danbooruタグ)

• A young girl in a school uniform, including a blazer and skirt, is standing alone in a park. She's wearing a tie and has a naughty grin, looking directly at the viewer with anime-like coloring. (LoRAなし、コントロール画像なし、英文)

• A young girl in a school uniform, including a blazer and skirt, is standing alone in a park. She's wearing a tie and has a naughty grin, looking directly at the viewer with anime-like coloring. (LoRA強度0.8、コントロール画像なし、英文)

• 1人の女の子が、学校制服を着て、公園で立っており、ニコニコと笑顔で、視聴者に悪戯っぽく見ています。アニメ調。 (LoRA強度0.8、コントロール画像なし、日本語文)

• 1girl, solo, standing, park, naughty, grin, looking at viewer, school uniform, blazer, jacket, shirt, skirt, tie, skirt lift by both hands, pink panties, panties, underwear, anime coloring (LoRA強度1.0、コントロール画像なし、danbooruタグ)

• A girl, dressed in a school uniform consisting of a blazer, jacket, shirt, and skirt, is standing alone in a park. She has a naughty grin and is looking directly at the viewer. The skirt is lifted by both hands, revealing pink panties underneath. The image is rendered in anime-style coloring. (LoRA強度1.0、コントロール画像なし、英文)

• コントロール画像

• 1girl, skirt lift by both hands, pussy, cleft of venus, uncensored (LoRA強度1.0、danbooruタグ, ワークフローなし)

• 1人の少女がスカートを両手で持ち上げ、まんこに露出している。 (LoRA強度1.0、日本語文、ワークフローなし)

(モザイクは手動で追加)

学習方法

musubi-tunerでvast.aiでRTX PRO 6000(0.713 $/hr, 232.2GB 19.2 CPU)で10エポック10810ステップ学習しました。4時間程度かかりました。

テキストエンコーダーに検閲解除版qwen2.5-vl-7b-instruct-abliteratedを使用しています。

musubi-tunerのインストール

https://github.com/kohya-ss/musubi-tuner/blob/main/README.ja.md

git clone https://github.com/kohya-ss/musubi-tuner/
cd musubi-tuner
pip install -e .
pip install schedulefree
mkdir models outputs
accelerate config

データセット

• dataset14
  • ネットから取ってきたdanbooruタグ付きNSFW画像1081枚

データセット

アスペクト比を保って1024x1024と同等の解像度で学習します。

設定はdataset.tomlに書きます。

[general]
resolution = [1024, 1024]
caption_extension = ".txt"
enable_bucket = true

[[datasets]]
image_directory = "dataset14"
control_directory = "dataset14-control"

モデル

musubi-tunerの文書に従い、次のモデルをmusubi-tunerフォルダのmodelsフォルダ以下に置きます。

• 拡散モデル https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/blob/main/split_files/diffusion_models/qwen_image_edit_2511_bf16.safetensors
• テキストエンコーダー(検閲解除版) https://huggingface.co/wangkanai/qwen2.5-vl-7b-instruct/tree/main
• VAE https://huggingface.co/Qwen/Qwen-Image/blob/main/vae/diffusion_pytorch_model.safetensors

wget -O models/qwen_image_edit_2511_bf16.safetensors https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/resolve/main/split_files/diffusion_models/qwen_image_edit_2511_bf16.safetensors
wget -O models/qwen2.5-vl-7b-instruct-abliterated.safetensors https://huggingface.co/wangkanai/qwen2.5-vl-7b-instruct/resolve/main/qwen2.5-vl-7b-instruct-abliterated.safetensors
wget -O models/diffusion_pytorch_model.safetensors https://huggingface.co/Qwen/Qwen-Image/resolve/main/vae/diffusion_pytorch_model.safetensors

画像キャッシュ作成

python src/musubi_tuner/qwen_image_cache_latents.py --dataset_config dataset.toml --vae  models/diffusion_pytorch_model.safetensors --model_version edit-2511

テキストエンコーダーキャッシュ作成

python src/musubi_tuner/qwen_image_cache_text_encoder_outputs.py --dataset_config dataset.toml  --text_encoder models/qwen2.5-vl-7b-instruct-abliterated.safetensors --batch_size 1 --model_version edit-2511

lora学習

• オプティマイザ―
  • RAdamScheduleFree
  • 学習率 1e-3
  • network_args ""weight_decay=5e-2", "betas=(0.9, 0.999)"
• network_dim 128
• エポック数 10 この設定で学習しました。

次のconfig.tomlに設定を書いています。

dit = "models/qwen_image_edit_2511_bf16.safetensors"
vae = "models/diffusion_pytorch_model.safetensors"
text_encoder = "models/qwen2.5-vl-7b-instruct-abliterated.safetensors"
dataset_config = "dataset.toml"
output_dir = "outputs"
output_name = "AnimeYuriNSFW"
sdpa = true
mixed_precision = "bf16"
timestep_sampling = "shift"
weighting_scheme = "none"
discrete_flow_shift = 2.2
optimizer_type = "schedulefree.RAdamScheduleFree"
optimizer_args = ["weight_decay=5e-2", "betas=(0.9, 0.999)"]
learning_rate = 1e-3
gradient_checkpointing = true
max_data_loader_n_workers = 2
persistent_data_loader_workers = true
network_module = "networks.lora_qwen_image"
network_dim = 128
max_train_epochs = 100
save_every_n_epochs = 10
seed = 42
model_version = "edit-2511"
compile = true

(100エポック学習する設定ですが10エポックで切り上げています)

次のコマンドでlora学習を行います。

accelerate launch --num_cpu_threads_per_process 1 src/musubi_tuner/qwen_image_train_network.py --config_file config.toml 

lora変換

最後にComfyUIで使える形式に変換します。

python src/musubi_tuner/convert_lora.py --input outputs/AnimeYuriNSFW.safetensors --output outputs/AnimeYuriNSFW-conv.safetensors --target other