Image tokenization

We introduce GIST, an adaptive image tokenizer whose token allocation is jointly controlled by pixel-level and semantic target errors, so that heavy compression can discard pixel detail while preserving what the image depicts.