Comment by filozopdasha

it actually sounds like a fun idea, but i have one question. do you think a lightweight CNN trained on synthetic candy layouts would outperform the deterministic decoder for messy real world photos?