it is time to train flux with anky, so that it can actually understand what anky is and how it looks like. right now the pipeline uses gemini, and as a consequence of that we are spending $ on each image genreation. on this computer there are two rtx4090s. we could use those to genreate the anky images. but for that, we need to come up with a traninig mechanism. and ffor that, i want to actually choose which of all the images that have been generated via anky apply. can you create an endpoint called /training that shows the images of anky a la tinder and allows me to say YES or NO for a given one (with a button below the image) so that we can do this? all the ones that i hit YES, move them to a training-images folder. not move them. copy them. then we need to rent a server on the cloud with more compute (training LORA flux on this machine runs out of memory) so that we can upload those images to that server, run the trianing, and then download the weights of that new model here on this machine and generate the images uisng that trained flux. thoughts? will this work? how would you do it practically? do we have enough images of anky or should we genreate more?