im particularly sympathetic to typo learning, which i think gets lost in the synthetic data discussion (mine here https://www.youtube.com/watch?v=yXPPcBlcF8U ht

im particularly sympathetic to typo learning, which i think gets lost in the synthetic data discussion (mine here https://www.youtube.com/watch?v=yXPPcBlcF8U https://www.youtube.com/watch?v=yXPPcBlcF8U ) but i think in this case you can still generate typos in images and it'd be learnable. not a hard issue relevant to the OP