STAR: STacked AutoRegressive Scheme for Unified Multimodal Learning

Benchmark Model Rank Results
image-editing-on-imgedit-dataSTAR-3B–Overall: 4.14Add: 4.26Adjust: 4.06Extract: 3.78Replace: 4.46…
image-editing-on-imgedit-dataSTAR-7B–Overall: 4.34Add: 4.33Adjust: 4.19Extract: 4.19Replace: 4.59…
image-generation-on-wiseSTAR-3B–Overall: 0.52Cultural: 0.58Time: 0.54Space: 0.48Biology: 0.49…
image-generation-on-wiseSTAR-7B–Overall: 0.66Cultural: 0.61Time: 0.67Space: 0.61Biology: 0.74…
text-to-image-generation-on-genevalSTAR-3B–Overall: 0.86
text-to-image-generation-on-genevalSTAR-7B–Overall: 0.91