kokiy365/Generative-Kanji

Generative-Kanji

Generative-Kanji is a fine-tuned Stable Diffusion model trained on 1,600 curated Kanji samples. The dataset was created from JIS X 208 characters, selectively choosing 16 radicals, and expanding into compound kanji. The goal is to explore novel character generation while retaining the visual and semantic logic of Kanji.

The model was fine-tuned using LoRA (Low-Rank Adaptation) on top of CompVis/stable-diffusion-v1-4, enabling the generation of novel kanji from English text prompts.

Training Details

• —Dataset: 1,600 samples (JIS X 208, 16 selected radicals)
• —Image resolution: 128×128
• —Captions: Multi-English phrases or single definition
• —Base model: CompVis/stable-diffusion-v1-4
• —Fine-tuning method: LoRA
• —LoRA Rank: 64
• —Alpha: 4
• —Epochs: 50
• —Batch size: 1
• —Learning rate: 1e-4
• —Scheduler: cosine