ellenjxu / stable-diff-kanji

Stable diffusion hallucinates new kanji

stable-diff-kanji

Follow instructions in CompVis stable diffusion to set up environment. Make sure you are in the proper conda env (conda activate ldm).

1. Run create_dataset.py to create dataset. Extracts Kanji images and corresponding English description from .xml (source).

data/
├── kanji/
│ ├── kanji_id1.jpg
│ ├── kanji_id2.jpg
│ ├── ...
│ └── descriptions.txt

2. Run ./train.sh
3. Generate images with python3 inference.py --prompt="prompt" --num_images=10