The purpose of this study is to improve the accuracy of automatic HTML generation from web page design images. pix2code is the state of art in this field. It is consist of design image learning part by CNN and HTML learning part by LSTM. We propose three improvements of adding a word embedding layer, applying VGG16 fine tuning to CNN, replacing LSTM to Bidirectional LSTM or GRU, and introducing attention mechanism. In the experiment, we employed a conventional data set which was used in pix2code and evaluated by a standard natural language generation metric called BLEU. As the results, the one of proposed models that contained the word embedding layer and the attention mechanism scored the accuracy of 99%. It overcomes the result of state of art scored 88%.