Preference Assessment on PreferImg SP 1.0 (seen)
96.6Assessment AccuracyPreferThinker
Evaluation Results
| Method | Links | |
|---|---|---|
| PreferThinker#Param=7B, Reference images as input for ICL=true2025.11 | 96.6 | |
| OpenAI-GPT-4oReference images as input for ICL=true2025.11 | 94.2 | |
| Claude 3.7Reference images as input for ICL=true2025.11 | 93.8 | |
| ViPer#Param=8B, Reference images as input for ICL=true2025.11 | 92.4 | |
| Doubao-1.5-vision-pro-32kReference images as input for ICL=true2025.11 | 92.2 | |
| GLM-4.1V-9B-Thinking#Param=9B, Reference images as input for ICL=true2025.11 | 86.8 | |
| Gemini-2.5-flashReference images as input for ICL=true2025.11 | 85.2 | |
| InternVL-3.5-8B#Param=8B, Reference images as input for ICL=true2025.11 | 80 | |
| Qwen2.5-VL-7B (Base)#Param=7B, Reference images as input for ICL=true2025.11 | 75.4 | |
| LLaVA-Reward#Param=8.2B, Reference images as input for ICL=false2025.11 | 53.2 | |
| ImageReward#Param=478M, Reference images as input for ICL=false2025.11 | 52.2 | |
| HPSv2#Param=2B, Reference images as input for ICL=false2025.11 | 51.4 | |
| Aesthetics#Param=304M, Reference images as input for ICL=false2025.11 | 50.6 | |
| CLIPScore#Param=428M, Reference images as input for ICL=false2025.11 | 50.2 | |
| PickScore#Param=986M, Reference images as input for ICL=false2025.11 | 49.6 | |
| UnifiedReward#Param=7B, Reference images as input for ICL=false2025.11 | 49 | |
| UnifiedReward-Think#Param=7B, Reference images as input for ICL=false2025.11 | 48.6 | |
| CycleReward#Param=477M, Reference images as input for ICL=false2025.11 | 48 |