Anyway, here's my comparison of giving each batch context from the previously generated pixels vs no context (64x64 pixels over 60 batch), goal is better coherence for each

Anyway, here's my comparison of giving each batch context from the previously generated pixels vs no context (64x64 pixels over 60 batch), goal is better coherence for each batch the context is like : "PLAN (your own earlier decisions about this picture): the main boundary between its two largest areas runs at about row 17 of 32, counting from the top. The focal object sits around columns 10-22, rows 6-18. Keep every pixel consistent with this."