bldng1337/atlas ? reverse-engineered prompt

Build me a Python project that lets me train and run a ControlNet style model for generating satellite terrain images and heightmaps with MESA.

I want a simple inference app where I can load a trained controlnet checkpoint, pick the device and precision, and generate results from terrain inputs. I also want the training side wired up so I can start training from a script, switch between different GPU friendly configs, and control the important settings from a config file or environment variables.

Please include the full dataset workflow too, so I can download the source data, filter it, merge imagery with elevation data, create feature maps, and add text prompts before training. If it helps, look up current docs online for any libraries or model details you need.

Keep it practical and runnable from the command line, with clear defaults, so someone can set it up and get to generating terrain imagery without a lot of manual work.