stable-diffusion-2-kanji-finetune

🧾 Model Card: Full Fine-Tuned – kentarrito/stable-diffusion-2-kanji-finetune

🈶 Stable Diffusion 2 – Kanji Fine-Tune (Full Model)

This is a full fine-tuned version of Stable Diffusion 2 on a custom dataset of kanji characters and their English meanings. The model was trained to generate kanji-style images based on English prompts such as "fire", "mountain", or "peace".

📦 Usage

from diffusers import StableDiffusionPipeline
import torch

pipe = StableDiffusionPipeline.from_pretrained(
    "kentarrito/stable-diffusion-2-kanji-finetune",
    torch_dtype=torch.float16
).to("cuda")

image = pipe(prompt="fire").images[0]
image.show()

🖼️ Generated Samples

See Github

🧠 Dataset

The dataset was built using:

• SVG files from KanjiVG
• English meanings from KANJIDIC2
• Uploaded to Hugging Face as kentarrito/kanji_dataset

Each training sample pairs an image of a kanji with one of its English meanings.

🎯 Limitations

• Generated images are kanji-like but may not be accurate or interpretable as real characters.
• The model may fail with abstract or multi-word prompts.

🧪 Training

• Training Code Github
• Model: stabilityai/stable-diffusion-2
• Fine-tuning: Full model training (UNet, text encoder, VAE)
• Framework: Hugging Face diffusers
• GPU: A40

📜 License

MIT License. Dataset sources are licensed under their respective terms.