There is other research that works with pixels of text, such as this recent paper I saw at COLM 2025 https://arxiv.org/abs/2504.02122.