It would be cool if it could continue from an existing song
For example, if I put in Moonlight Sonata, it would continue in that style.
Yes very good point ! That is something we are working on indeed! But not sure if it will be integrated in this app soon.
@misnaej Perhaps it would be easier to do if, instead of processing the song into MIDI yourself, the option you’d leave users is to paste MIDI they’ve already processed themselves. That way, you are not tacking on a bunch a time to how long it takes to generate outputs. Then, you could also introduce a character limit, since GPT-2 takes longer, the longer the input. That also avoids unduly long waiting times.
These are very good points. Would it speed up the process to input the processed midi? I am not sure because encoding of one file goes pretty fast. But I will check. Another problem I see is that the midi needs to be first encoded with our encoding (which is not a big deal but might be not user friendly). About the size, for sure it is a limit, and an idea would be to (randomly?) select one part of the uploaded midi to use as a prompt, which should be quite short indeed, maybe one or two tracks (determined based on a character limit as you suggest). Working on this now. But it will take a little while at this time of year ;-)