You could have designed state of the art positional encoding

There is this really interesting blog post about making rope (by the main author of the paper) multimodal as used by qwen2 vl. it's in chinese but google translate does a pretty good job: https://spaces.ac.cn/archives/10040