Research map / Image generation / Autoregressive and latent generation
Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Compare2Score trains an mPLUG-Owl2-based LMM to output five-level pairwise quality comparisons built from MOS differences within each IQA dataset, then converts test-time soft comparison probabilities against anchor images into a single quality score via MAP estimation under Thurstone's Case V.
Problem
Existing LMM-based NR-IQA models trained on absolute ratings cannot be reliably combined across IQA datasets because each dataset has different perceptual scales, and comparison-based methods are impractical to convert into individual scores for real-world use.
Approach
Generate comparative instruction-response pairs by comparing MOSs within the same dataset (five levels via empirical significance thresholds), fine-tune an LMM as a visual quality comparator, and at inference compute a probability matrix against anchor images via weighted softmax over the five level tokens, then estimate the quality score by MAP under Thurstone's Case V.
Key results
- Compare2Score achieves SRCC 0.972 on LIVE, 0.950 on CSIQ, 0.952 on KADID-10k, and 0.919 on BID under the intra-dataset setting, outperforming LIQE and Q-Align.
- In cross-dataset tests, Compare2Score reaches SRCC 0.823 on TID2013, 0.906 on SPAQ, and 0.730 on AGIQA-3K, the best among compared methods.
- The probability-matrix soft comparison improves SRCC over count-matrix inference for mPLUG-Owl2 and Co-Instruct and other zero-shot LMMs, e.g. mPLUG-Owl2 on BID rises from 0.217 to 0.551.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the probability matrix instead of the count matrix (ablation) | Compare2Score · LIVE | SRCC: 0.888 -> 0.974 | Helped |
| Use the probability matrix instead of the count matrix (ablation) | Compare2Score · CSIQ | SRCC: 0.875 -> 0.942 | Helped |
| Use the probability matrix instead of the count matrix (ablation) | Compare2Score · BID | SRCC: 0.778 -> 0.921 | Helped |
| Use the probability matrix instead of the count matrix (ablation) | Compare2Score · CLIVE | SRCC: 0.816 -> 0.934 | Helped |
| Select anchors by low MOS variability rather than randomly (ablation) | Compare2Score · LIVE | SRCC: 0.954 -> 0.972 | Helped |
| Select anchors by low MOS variability rather than randomly (ablation) | Compare2Score · CSIQ | SRCC: 0.939 -> 0.950 | Helped |
| Select anchors by low MOS variability rather than randomly (ablation) | Compare2Score · KADID-10k | SRCC: 0.944 -> 0.952 | Helped |
| Select anchors by low MOS variability rather than randomly (ablation) | Compare2Score · BID | SRCC: 0.881 -> 0.919 | Helped |
| Select anchors by low MOS variability rather than randomly (ablation) | Compare2Score · CLIVE | SRCC: 0.890 -> 0.914 | Helped |
| Select anchors by low MOS variability rather than randomly (ablation) | Compare2Score · KonIQ-10k | SRCC: 0.915 -> 0.931 | Helped |
| Increase anchors per quality interval beyond 1 (ablation) | Compare2Score · evaluated IQA datasets | SRCC | No clear effect |
| Use anchors from KADID-10k instead of KonIQ-10k (ablation) | Compare2Score · KADID-10k | SRCC: 0.952 -> 0.957 | Helped |
Surprising findings
- Using only one anchor image per quality interval (beta=1) is sufficient; increasing the number of anchors does not improve performance.
- Random anchor selection performs competitively with the proposed variance-based selection, remaining close to the state of the art.
Limitations
- Paired comparison requires several anchor comparisons per test image, adding inference cost compared with direct scoring.
- Performance on some datasets such as CSIQ with zero-shot LMMs remains low, and the method depends on MOS availability in training datasets.
Experimental setup
Datasets: LIVE CSIQ KADID-10k BID CLIVE KonIQ-10k TID2013 SPAQ AGIQA-3K
Benchmarks: LIVE CSIQ KADID-10k BID CLIVE KonIQ-10k TID2013 SPAQ AGIQA-3K
Metrics: SRCC PLCC prediction accuracy
Baselines: NIQE ILNIQE Ma19 PaQ2PiQ KonCept MUSIQ DBCNN HyperIQA TreS UNIQUE LIQE Q-Align
Builds on: mPLUG-Owl2 Q-Align LIQE Thurstone's Case V model
Closest papers by meaning
- Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption (CVPR 2025 · similarity 0.91)
- Refine-IQA: Multi-Stage Reinforcement Finetuning for Perceptual Image Quality Assessment (AAAI 2026 · similarity 0.89)
- IQA-Spider: Unifying Multi-Granularity Image Quality Assessment with Reasoning, Grounding and Referring (ICML 2026 · similarity 0.89)
- Augmenting Perceptual Super-Resolution via Image Quality Predictors (CVPR 2025 · similarity 0.89)
- Revisiting MLLM Based Image Quality Assessment: Errors and Remedy (AAAI 2026 · similarity 0.89)
- VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.