Generation Enhances Understanding in Unified Multimodal Models via Multi-Representation Generation

Benchmark Model Rank Results
visual-question-answering-vqa-on-hallusionbench-average-accuracyOpenUni-3.6B + UniMRG#9Accuracy: 64.56
visual-question-answering-vqa-on-realworldqaOpenUni-3.6B + UniMRG#71Accuracy: 66.01
text-to-image-generation-on-genevalOpenUni-3.6B + UniMRG#155Overall: 0.558