Image Quality Assessment on PIPAL
69SRCCDeQA-Score
Evaluation results
| Method | Links | ||
|---|---|---|---|
| DeQA-ScoreTraining Dataset=KonIQ, SPAQ, KADID, PIPAL2026.01 | 69 | 72.4 | |
| Q-HawkeyeTraining Dataset=KonIQ2026.01 | 55 | 55.1 | |
| Q-HawkeyeMethodology=MLLM-Based Methods2026.01 | 55 | 55.1 | |
| TATARCategory=Unified2026.03 | 50.1 | 49.8 | |
| DeQA-ScoreTraining Dataset=KonIQ, SPAQ, KADID2026.01 | 49.6 | 49.5 | |
| VisualQuality-R1Venue=NeurIPS’25, Methodology=MLLM-Based Methods2026.01 | 48.6 | 53.1 | |
| Q-InsightCategory=Specialized, Evaluation Domain=Cross-domain, Evaluation Protocol=Single-task2026.03 | 48.6 | 47.4 | |
| DeQA-ScoreVenue=CVPR’25, Methodology=MLLM-Based Methods2026.01 | 47.8 | 47.2 | |
| DeQA-ScoreTraining Dataset=KonIQ, SPAQ2026.01 | 47.4 | 46.8 | |
| Q-insightTraining Dataset=KonIQ, KADIS2026.01 | 47.4 | 48.6 | |
| DeQACategory=Specialized, Evaluation Domain=Cross-domain, Evaluation Protocol=Single-task2026.03 | 47.2 | 47.8 | |
| UniPerceptCategory=Unified, Training Protocol=Retrained under our protocol2026.03 | 45.9 | 46.8 | |
| Q-insightVenue=NeurIPS’25, Methodology=MLLM-Based Methods2026.01 | 45.3 | 46.1 | |
| ManIQAVenue=CVPR’22, Methodology=Non-MLLM Deep-Learning2026.01 | 45.2 | 45.7 | |
| InternVL3.5-38BCategory=Open-Source, Evaluation Domain=Cross-domain2026.03 | 44.8 | 45.7 | |
| Qwen-SFTVenue=arXiv’25, Methodology=MLLM-Based Methods2026.01 | 44.2 | 47.3 | |
| VisualQuality-R1Training Dataset=KADID, SPAQ2026.01 | 43.4 | 43.7 | |
| MUSIQVenue=ICCV’21, Methodology=Non-MLLM Deep-Learning2026.01 | 43.1 | 43.1 | |
| CLIP-IQA+Venue=AAAI’23, Methodology=Non-MLLM Deep-Learning2026.01 | 43.1 | 43.7 | |
| Q-AlignVenue=ICML’24, Methodology=MLLM-Based Methods2026.01 | 41.9 | 40.3 | |
| LLaVA-1.5-8BCategory=Open-Source, Evaluation Domain=Cross-domain2026.03 | 41.7 | 40.7 | |
| InternVL3-78BCategory=Open-Source, Evaluation Domain=Cross-domain2026.03 | 41.5 | 45.7 | |
| QwenVL-3-32BCategory=Open-Source, Evaluation Domain=Cross-domain2026.03 | 41.4 | 40.2 | |
| HyperIQAVenue=CVPR’20, Methodology=Non-MLLM Deep-Learning2026.01 | 40.3 | 41 | |
| Q-AlignCategory=Specialized, Evaluation Domain=Cross-domain, Source=Cited from UniPercept2026.03 | 40.3 | 41.9 | |
| NIMAVenue=TIP’18, Methodology=Non-MLLM Deep-Learning2026.01 | 39.9 | 39 | |
| QwenVL-2.5-72BCategory=Open-Source, Evaluation Domain=Cross-domain2026.03 | 38.1 | 40.7 | |
| DBCNNVenue=ICSIPA’19, Methodology=Non-MLLM Deep-Learning2026.01 | 37.1 | 38.4 | |
| Q-InsightCategory=Specialized, Evaluation Domain=Cross-domain, Source=Cited from UniPercept2026.03 | 36.9 | 36.8 | |
| Compare2ScoreVenue=NIPS’24, Methodology=MLLM-Based Methods2026.01 | 34.2 | 35.4 | |
| GPT-4oCategory=Proprietary, Evaluation Domain=Cross-domain2026.03 | 32.5 | 34.9 | |
| BRISQUEVenue=TIP’12, Methodology=Handcrafted2026.01 | 23.2 | 26.7 | |
| Gemini-2.5-ProCategory=Proprietary, Evaluation Domain=Cross-domain2026.03 | 22.5 | -1.9 | |
| NIQEVenue=SPL’12, Methodology=Handcrafted2026.01 | 16.1 | 19.5 | |
| GLM-4.5Category=Open-Source, Evaluation Domain=Cross-domain2026.03 | 1.3 | 2 | |
| TOPIQ-FR(TIP24)Feature Domain=Feature comparison, trained on dataset=PIPAL2026.04 | 0.809 | 0.837 | |
| DBIQA(TPAMI25)Feature Domain=Feature comparison, trained on dataset=PIPAL2026.04 | 0.769 | 0.745 | |
| Causal Disentanglement (Ours)Feature Domain=Causal Disentanglement, fine-tuning=true2026.04 | 0.713 | 0.727 | |
| PieAPP(CVPR18)Feature Domain=Feature comparison2026.04 | 0.701 | 0.702 | |
| DeepDC(TIP25)Feature Domain=Feature comparison2026.04 | 0.67 | 0.669 | |
| Causal Disentanglement (Ours w/o f)Feature Domain=Causal Disentanglement2026.04 | 0.666 | 0.684 | |
| DeepCausal(CVPR25)Feature Domain=Feature comparison2026.04 | 0.657 | 0.675 | |
| ARNIQAOpinion-Unaware=false2024.03 | 0.634 | 0.666 | |
| ADISTSFeature Domain=Feature comparison2026.04 | 0.631 | 0.645 | |
| DISTS(TPAMI20)Feature Domain=Feature comparison2026.04 | 0.627 | 0.645 | |
| DSDFeature Domain=Feature comparison2026.04 | 0.624 | 0.621 | |
| DeepJSD(TIP25)Feature Domain=Feature comparison2026.04 | 0.598 | 0.612 | |
| QualiCLIP+Opinion-Unaware=false2024.03 | 0.586 | 0.59 | |
| CONTRIQUEOpinion-Unaware=false2024.03 | 0.581 | 0.611 | |
| UniPerceptModel Category=Ours2025.12 | 0.581 | 0.594 | |
| LPIPS(CVPR18)Feature Domain=Feature comparison2026.04 | 0.573 | 0.611 | |
| GMSDFeature Domain=Traditional2026.04 | 0.569 | 0.614 | |
| Re-IQAOpinion-Unaware=false2024.03 | 0.568 | 0.587 | |
| GRepQOpinion-Unaware=false2024.03 | 0.554 | 0.568 | |
| PDLFeature Domain=Feature comparison2026.04 | 0.554 | 0.564 | |
| CLIP-IQA+Opinion-Unaware=false2024.03 | 0.552 | 0.558 | |
| MS-SSIMFeature Domain=Traditional2026.04 | 0.545 | 0.571 | |
| VIFFeature Domain=Traditional2026.04 | 0.545 | 0.572 | |
| RALICategory=Non-MLLM Deep-learning, Training Dataset=KonIQ2025.10 | 0.528 | 0.527 | |
| VSIFeature Domain=Traditional2026.04 | 0.526 | 0.548 | |
| Q-DeepSightCategory=MLLM (w/ reasoning), Training Dataset=KonIQ2026.04 | 0.502 | 0.51 | |
| LIQECategory=Non-MLLM Deep-learning, Training Dataset=KonIQ2025.10 | 0.501 | 0.48 | |
| DeepWSD(MM22)Feature Domain=Feature comparison2026.04 | 0.5 | 0.503 | |
| Q-DeepSight†Category=MLLM (w/ reasoning), Training Dataset=KonIQ + DQ-7K2026.04 | 0.496 | 0.501 | |
| Q-InsightModel Category=Specialized, Domain=Cross domain2025.12 | 0.486 | 0.474 | |
| SSIMFeature Domain=Traditional2026.04 | 0.486 | 0.489 | |
| Tool-IQATraining Dataset=KonIQ2026.06 | 0.479 | 0.483 | |
| DeQACategory=VLM (w/o & w/ reasoning), Trained on KonIQ=true2026.01 | 0.478 | 0.472 | |
| DeQACategory=MLLM-based, Training Dataset=KonIQ2025.10 | 0.478 | 0.472 | |
| DeQA-ScoreCategory=MLLM (w/o reasoning), Training Dataset=KonIQ2026.04 | 0.478 | 0.472 | |
| Q-ProbeCategory=MLLMs-based2026.01 | 0.474 | 0.476 | |
| Q-InsightCategory=MLLM-based, Training Dataset=KonIQ2025.10 | 0.474 | 0.486 | |
| Q-Insight†Category=MLLM (w/ reasoning), Training Dataset=KonIQ + DQ-7K2026.04 | 0.474 | 0.486 | |
| DeQAModel Category=Specialized, Domain=Cross domain2025.12 | 0.472 | 0.478 | |
| VisualQuality-R1Category=MLLMs-based2026.01 | 0.469 | 0.458 | |
| Zoom-IQACategory=VLM (w/o & w/ reasoning), Trained on KonIQ=true2026.01 | 0.465 | 0.468 | |
| Zoom-IQACategory=MLLM (w/ reasoning), Training Dataset=KonIQ2026.04 | 0.465 | 0.468 | |
| Tool-augmented Zoom-IQATraining Dataset=KonIQ2026.06 | 0.465 | 0.468 | |
| NLPDFeature Domain=Traditional2026.04 | 0.464 | 0.489 | |
| ManIQACategory=Non-MLLM Deep-learning, Training Dataset=KonIQ2025.10 | 0.457 | 0.419 | |
| ManIQACategory=Non-VLM, Trained on KonIQ=true2026.01 | 0.452 | 0.457 | |
| CLIP-IQA+Category=Non-MLLM Deep-learning, Training Dataset=KonIQ2025.10 | 0.452 | 0.427 | |
| ManIQACategory=Non-MLLM, Training Dataset=KonIQ2026.04 | 0.452 | 0.457 | |
| GRepQOpinion-Unaware=false, Training Dataset=FLIVE2024.03 | 0.448 | 0.46 | |
| InternVL3.5-38BModel Category=Open-Source, Domain=Cross domain testing2025.12 | 0.448 | 0.457 | |
| LIQEOpinion-Unaware=false, training_dataset=CLIVE2024.03 | 0.446 | 0.445 | |
| Qwen-SFTCategory=MLLM (w/ reasoning), Training Dataset=KonIQ2026.04 | 0.442 | 0.473 | |
| VisualQuality-R1Category=VLM (w/o & w/ reasoning), Trained on KonIQ=true2026.01 | 0.441 | 0.451 | |
| VisualQuality-R1Category=MLLM (w/ reasoning), Training Dataset=KonIQ2026.04 | 0.441 | 0.451 | |
| VisualQuality-R1Training Dataset=KonIQ2026.06 | 0.441 | 0.451 | |
| VisualQuality-R1Category=MLLM-based, Training Dataset=KonIQ2025.10 | 0.438 | 0.441 | |
| Q-InsightCategory=VLM (w/o & w/ reasoning), Trained on KonIQ=true2026.01 | 0.435 | 0.458 | |
| InternVL3-8BModel Category=Open-Source, Domain=Cross domain testing2025.12 | 0.435 | 0.459 | |
| Q-InsightCategory=MLLM (w/ reasoning), Training Dataset=KonIQ2026.04 | 0.435 | 0.458 | |
| QwenVL-3-Instruct-8BModel Category=Open-Source, Domain=Cross domain testing2025.12 | 0.434 | 0.427 | |
| MUSIQCategory=Non-VLM, Trained on KonIQ=true2026.01 | 0.431 | 0.431 | |
| MUSIQCategory=Non-MLLM Deep-learning, Training Dataset=KonIQ2025.10 | 0.431 | 0.431 | |
| MUSIQCategory=Non-MLLM, Training Dataset=KonIQ2026.04 | 0.431 | 0.431 | |
| Q-InsightCategory=MLLMs-based2026.01 | 0.429 | 0.462 | |
| CLIP-IQA+Category=VLM (w/o & w/ reasoning), Trained on KonIQ=true2026.01 | 0.419 | 0.427 |