The situation I encountered here is that I have two inputs(for instance, image embedding, etc.) into the first lstm of a series of lstms to predict the next word to generate sentence(from the second lstm, it started to predict the next word from the current input word). The length of each of the two inputs is 512. Merely for the first input, it increases the measurement, say, for instance, perplexity, by about 3 from no this input at all. Merely for the second input, it increases the measurement, say, for instance, perplexity, by about 1 from no input at all. The problem is: Is it possible to combine these two inputs into a model that can produce a result of increasement more than 3 or the larger amount of increasement of the former two inputs models? If it is, how to build a model or what model should I build to combine them to do so?