Instruction-guided image editing preference prediction on GenAI-Bench
67.5AccuracyREWARDHARNESS (Qwen)
Evaluation results
| Method | Links | |
|---|---|---|
| REWARDHARNESS (Qwen)Sub-Agent=Qwen2026.05 | 67.5 | |
| EDITREWARDBackbone=MiMo-VL-7B-SFT2025.09 | 65.72 | |
| EditReward (MiMo)Model Type=Open-Source Models, Backbone=MiMo2026.05 | 65.7 | |
| REWARDHARNESS (Gemini-2.0-Flash)Sub-Agent=Gemini-2.0-Flash2026.05 | 64.4 | |
| EditReward (Qwen)Model Type=Open-Source Models, Backbone=Qwen2026.05 | 64 | |
| EDITREWARDBackbone=Qwen2.5-VL-7B2025.09 | 63.97 | |
| ADIEEMethod Category=Open-Source VLMs2025.09 | 59.96 | |
| GPT-5Method Category=Proprietary Models2025.09 | 59.61 | |
| GPT-5Model Type=Proprietary Models2026.05 | 59.6 | |
| MiMo-VL-7BModel Type=Open-Source Models2026.05 | 57.9 | |
| MiMo-VL-7B-SFT-2508Method Category=Open-Source VLMs2025.09 | 57.89 | |
| Gemini-2.5-FlashMethod Category=Proprietary Models2025.09 | 57.01 | |
| Gemini-2.5-FlashModel Type=Proprietary Models2026.05 | 57 | |
| GPT-4oMethod Category=Proprietary Models2025.09 | 53.54 | |
| GPT-4oModel Type=Proprietary Models2026.05 | 53.5 | |
| Gemini-2.0-FlashMethod Category=Proprietary Models2025.09 | 53.32 | |
| Gemini-2.0-FlashModel Type=Proprietary Models2026.05 | 53.3 | |
| Claude-Haiku-4.5Model Type=Proprietary Models2026.05 | 47.1 | |
| Qwen2.5-VL-3B-InstMethod Category=Open-Source VLMs2025.09 | 42.76 | |
| Qwen2.5-VL-7BModel Type=Open-Source Models2026.05 | 40.5 | |
| Qwen2.5-VL-7B-InstMethod Category=Open-Source VLMs2025.09 | 40.48 | |
| Qwen2.5-VL-32BModel Type=Open-Source Models2026.05 | 39.3 | |
| Qwen2.5-VL-32B-InstMethod Category=Open-Source VLMs2025.09 | 39.28 | |
| RandomMethod Category=Baseline2025.09 | 25.9 |