What kind of compression is used? (the blog post says model is trained on compressed frames) Is there an architecture diagram of the model and/or a minGPT-styl

What kind of compression is used? (the blog post says model is trained on compressed frames) Is there an architecture diagram of the model and/or a minGPT-style implementation?