I made a bunch of screenshots of the game, scaled them down 2x, because in original 1080p screenshots 1 pixel was made of 4x4 pixels, so it took out some redundancy, then I cut them into 512x512 manually to get as many features from each screenshot, then I used civitai lora trainer, i made brief captions for every picture and run the training with standard settings. If you wanna train something similar definitely use lossless screenshots because pictures you find online have a lot of compression on them and it'll fudge up the model.