Diffusion-KanjiGrade1

Diffusion-KanjiGrade1 is a fine-tuned Stable Diffusion model trained on 80 curated Kanji samples.
The dataset was derived from KANJIDIC2, filtered to include Grade 1 characters with stroke counts <10 and representing the top 0.05 most frequent words.
Images were rendered at 128×128 resolution, with captions containing only the English meaning(s) of each Kanji.
Training prompts included both single-word and multi-word meanings.

Training Configuration

• Base model: CompVis/stable-diffusion-v1-4
• LoRA rank: 32
• Alpha: 4
• Batch size: 1
• Learning rate: 1.0e-04
• LR scheduler: cosine
• Checkpointing steps: 3000
• Total training steps: 3000

Sample Generation

prompt = "river"

Prompt

Screenshot