Research map / Vision-language and multimodal / Visual question answering
Q-SiT: Teaching LMMs for Image Quality Scoring and Interpreting
Zicheng Zhang, Haoning Wu, Ziheng Jia, Weisi Lin, Guangtao Zhai
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 D1 scoring data (IQA datasets in QA form) (ablation) | Q-SiT · scoring (avg SRCC/PLCC) | score: 0.800 -> 0.861 | Improved |
| add D1 scoring data (ablation) | Q-SiT · interpreting (LLVisionQA) | accuracy: 0.754 -> 0.756 | Mixed |
| add D2 low-level interpreting data (ablation) | Q-SiT · interpreting (LLVisionQA) | accuracy: 0.721 -> 0.756 | Improved |
| add D3 extended conversations (ablation) | Q-SiT · interpreting (LLVisionQA) | accuracy: 0.744 -> 0.756 | Improved |
| add D3 extended conversations (ablation) | Q-SiT · scoring (avg SRCC/PLCC) | score: 0.859 -> 0.861 | No clear effect |
| add coarse-grained data mix ratio prediction (ablation) | Q-SiT · interpreting (LLVisionQA) | accuracy: 0.732 -> 0.752 | Improved |
| add coarse-grained data mix ratio prediction (ablation) | Q-SiT · scoring (avg SRCC/PLCC) | score: 0.855 -> 0.859 | Improved |
| add fine-grained loss-based ratio adjustment on top of CGR (ablation) | Q-SiT · interpreting (LLVisionQA) | accuracy: 0.752 -> 0.756 | Improved |
| add fine-grained loss-based ratio adjustment on top of CGR (ablation) | Q-SiT · scoring (avg SRCC/PLCC) | score: 0.859 -> 0.861 | Improved |
| predict mix ratio with 0.5B model instead of searching on 7B | Q-SiT · training cost | A800 GPU hours: 2,800 -> 72 | Improved |
| use Q-SiT scoring-plus-interpreting training instead of interpreting-only Q-Instruct | Q-SiT 7B · LLVisionQA | overall accuracy: 72.30% -> 75.65% | Improved |
| shrink Q-SiT to 0.5B (Q-SiT-mini) | Q-SiT · inference | latency, VRAM, FLOPs: 49% latency reduction, 25% VRAM, 92% fewer FLOPs | Improved |
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.
- Q-Align — Q-Align: Teaching LMMs for Visual Scoring via Discrete Text-Defined Levels (ICML 2024) · better on 6 of 7 shared results
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
- Q-Align: Teaching LMMs for Visual Scoring via Discrete Text-Defined Levels (ICML 2024 · similarity 0.90)
- Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.90)
- Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution (CVPR 2025 · similarity 0.89)
- IQA-Spider: Unifying Multi-Granularity Image Quality Assessment with Reasoning, Grounding and Referring (ICML 2026 · similarity 0.88)
- Grounding-IQA: Grounding Multimodal Language Model for Image Quality Assessment (ICLR 2026 · similarity 0.88)
- Revisiting MLLM Based Image Quality Assessment: Errors and Remedy (AAAI 2026 · similarity 0.87)
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.
