Dataset for longer levels #96
Description
Both because it could result in better diffusion models, and because it might give us useful data to train MarioGPT, let's create a new dataset with longer levels. Currently, the levels are sliced into 16 by 16 segments and the diffusion models are trained on that. It should be possible to train it on 16 by 128 samples. Look at Mar1and2-data.bat and see which commands it runs, and then see if those commands can be run to instead generate longer level samples, and also see what kinds of captions get assigned to these longer levels. It may no longer be appropriate to have a stride of 1. Make the dataset for longer levels, and then see if it is possible to train a diffusion model on it.
Start with an unconditional model first for simplicity (captions not relevant), and then try a conditional one.