ResearchBenchmarksEfficiency Comparison on CHAIR benchmarkFollow3.36Avg Time (s/caption)Baseline2.109610.549818.9927.4302Feb 4, 2026Evaluation ResultsMethodMethodLinksAvg Time (s/caption)Peak Memory (MB)Latency (ms/token)Throughput (token/s)BaselineBackbone=LLaVA-1.5Backbone=LLaVA-1.52026.023.3614,629.2131.2432.06KVSmoothBackbone=LLaVA-1.5Backbone=LLaVA-1.52026.023.6114,625.0634.3329.18SPARCBackbone=LLaVA-1.5Backbone=LLaVA-1.52026.024.1314,644.6135.228.44PAIBackbone=LLaVA-1.5Backbone=LLaVA-1.52026.026.6815,003.2556.4417.73OPERABackbone=LLaVA-1.5Backbone=LLaVA-1.52026.0234.6219,531.97313.783.43