SPAR: Semantic-Pixel Self-Alignment and Adaptive Routing for Unified Multimodal Models

Benchmark Model Rank Results
image-editing-on-imgedit-dataSPAR-3B–Overall: 4.01Add: 4.31Adjust: 3.93Extract: 2.32Replace: 4.52…
image-generation-on-wiseSPAR-3B (extra data)–Overall: 0.64
text-to-image-generation-on-genevalSPAR-3B (extra data)–Overall: 0.91
text-to-image-generation-on-wiseSPAR-1B–Overall: 0.57Cultural: 0.54Biology: 0.51
text-to-image-generation-on-wiseSPAR-3B–Overall: 0.64Cultural: 0.67Biology: 0.61
visual-question-answering-on-mm-vetSPAR-1B–GPT-4 score: 59.8
visual-question-answering-on-mm-vetSPAR-3B–GPT-4 score: 62.2