Bad Result when trying to train model with input is images

[qbKhanh]

This publication is so amazing. I get a very good result when use generating image while training progress (generate_indigo_image) as in your source code but it takes much time when train with large dataset.

So I separated the generating image progress out of training progress. I used with same setup to gen an image dataset then I pass this dataset to train model. But the result is very very bad. The prediction string is so long and look stupid, and the evaluation metrics are too.

I tried to visualize the image and corresponding label, nothing is wrong.

So I raise this issue hope that can hear from you some ideas or solutions to solve this problem!
Thanks for spending time to read my issue!