Concept

1. Idea: snake game without engine and only with transformers
2. Hot it works: write snake core game, write Neural Network that can play it, train it with RL
3. During the training and playing, store history of key inputs, screenshots that will be used as context for the next screenshot
4. Then train diffusion model to generate screenshot base on history of key inputs and screenshots
5. Use this model to generate new screenshots based on the previous screenshot and key inputs
6. We measure LPIPS (Zhang et al., 2018) and PSNR (Wang et al., 2004) between the original and generated screenshots to evaluate the quality of the generated screenshots