This is the result of me being bored at 3 AM during exam week when I can't sleep because REM cycles or something.
Anyway, this is just a rough sketch of how captions can be generated from just an image.
First, the image gets processed through Clarifai and returns tags that describe the image. I then took hunkim's word-RNN repository to take a set of encouraging messages and generate a heuristic of a caption based on the relevant captions.
Once the RNN spits out a garbage caption, I combine that with a Markov chain trained on the original messages so that the grammar is not too bad, and there is slight relevancy to the original image.
Install the Clarifai Python client. Set it up according to their instructions, including getting your own Client ID and Secret.
Install the remaining dependencies, including markovify, numpy, and tensorflow if you don't have them already.
Change line 33 of index.py to the URL of an image you want to caption.
Run train.py to train the RNN, then run index.py to test it out!