Regional Attention

Cross-attend to different prompt based on region of image

Bitmask

• illustration of ice dragon fighting in a fantasy battleground
• illustration of fire dragon fighting in a fantasy battleground
• illustration of sakura tree in full bloom, petals falling, spring day, calm sky, rolling hills, grass, flowers masterpiece, dramatic, highly detailed, high dynamic range

Generated via regional_attn.py, at commit bddf6de

Horizontal split

• macaron on acacia wood platter
• illustration of sakura tree in full bloom, petals falling, spring day, calm sky, rolling hills, grass, flowers masterpiece, dramatic, highly detailed, high dynamic range
• illustration of ice dragon girl, wings, winter day, masterpiece, dramatic, highly detailed, high dynamic range, watercolor (medium)
• illustration of fire dragon girl, wings, night, masterpiece, dramatic, highly detailed, high dynamic range, aurora borealis

Generated via regional_attn.py, at commit a10c207

Setup

Linux/CUDA:

python3 -m venv venv
. venv/bin/activate
pip install wheel
pip install -r requirements.txt
# optional dep for playing around with high-pass noise interpolation
pip install dctorch --no-deps

Run

Ensure virtualenv has been activated.

python -m script.play