Mapping screenshots to code is not hard. By having the model simply memorize the screenshots to code mappings of the training data can give you almost 100% accu

Mapping screenshots to code is not hard. By having the model simply memorize the screenshots to code mappings of the training data can give you almost 100% accuracy (for some demo). What is hard is if given a new screenshot, how would this model generalize. To have something work for mobiles is a much easier task than having something work for other more complex UI though. Looking forward to seeing more updates on this!