This is my default explanation for visual impairments in LLMs, they're trying to compress the image into about 3000 tokens, you're going to lose a lot in the name of efficiency.