Here are some examples that demonstrate the performance of the model. I didn't just want to optimize the metrics, that can be found in the literature if that means images found in the wild will not be recognized.

Where means the code generates the same image, means that the predicted code compiles and the image is close to the original. means that there are significant differences between the original and the predicted code.

	BLEU	normed edit distance	token accuracy
Performance	0.88	0.10	0.60