Research map / Vision-language and multimodal / Multimodal LLMs
Towards Open-ended Visual Quality Comparison
Haoning Wu, Hanwei Zhu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Annan Wang, Wenxiu Sun, Qiong Yan, Xiaohong Liu, Guangtao Zhai and 2 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-image comparative data (Merge2Compare and Teach2Compare) to single-image Q-Instruct training (ablation) | mPLUG-Owl2 · Q-Bench PAIR-A1 | accuracy: 53.15 -> 80.18 | Improved |
| add comparative data (ablation) | mPLUG-Owl2 · MICBench | accuracy: 45.98 -> 83.33 | Improved |
| add Merge2Compare to Q-Instruct (ablation) | mPLUG-Owl2 · Q-Bench PAIR-A1 | accuracy: 53.15 -> 68.67 | Improved |
| train with only GPT-4V labels (Teach2Compare) and no Merge2Compare (ablation) | mPLUG-Owl2 · Q-Bench PAIR-A1 | accuracy: 78.28 vs GPT-4V 78.07 | No clear effect |
| combine hetero-sourced Merge2Compare with Teach2Compare (full) (ablation) | mPLUG-Owl2 · MICBench | accuracy: 76.41 -> 83.33 | Improved |
| use image-text interleaved format with ordinal labels instead of piling images (ablation) | Co-Instruct · MICBench | accuracy: 78.92 -> 83.33 | Improved |
| use learnable special image start/end tokens instead of ordinal text (ablation) | Co-Instruct · MICBench | accuracy: 83.33 -> 80.12 | Hurt |
| add top-similarity pair removal to Merge2Compare construction (ablation) | Merge2Compare data · human check of 250 samples | correctness rate: 72% -> 96% | Improved |
| add comparative training data on single-image evaluation (ablation) | mPLUG-Owl2 · Q-Bench SINGLE-A1 | accuracy: 73.38 -> 77.12 | Improved |
| Co-Instruct versus GPT-4V on detailed reasoning score | Co-Instruct · Q-Bench PAIR-A2 | summed GPT score: 4.89 -> 4.82 | Hurt |
| extend context window for LMMs lacking visual token reduction on four images | LLaVA, InternLM-XComposer2 · MICBench | accuracy: 55.70 (three) -> 46.51 (four) for XComposer2 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- mPLUG-Owl2 — mPLUG-Owl2: Revolutionizing Multi-modal Large Language Model with Modality Collaboration (CVPR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.92)
- Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption (CVPR 2025 · similarity 0.87)
- R4-CGQA: Retrieval-based Vision Language Models for Computer Graphics Image Quality Assessment (CVPR 2026 · similarity 0.87)
- IQA-Spider: Unifying Multi-Granularity Image Quality Assessment with Reasoning, Grounding and Referring (ICML 2026 · similarity 0.87)
- Evaluation of Visual Processing Should Be Human-Centered, Not Metric-Centered (NeurIPS 2026 · similarity 0.87)
- VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank (NeurIPS 2025 · similarity 0.86)
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.
