UniME-V2: MLLM-as-a-Judge for Universal Multimodal Embedding Learning