Research map / Vision-language and multimodal / Multimodal LLMs

IQA-Spider: Unifying Multi-Granularity Image Quality Assessment with Reasoning, Grounding and Referring

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 0 citations (OpenAlex)

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.

Proposes IQA-Spider, a unified LMM-based IQA framework covering global/local description, pixel-level grounding and region referring, with a new 33K-sample dataset and a training-free text-to-point grounding paradigm that maps positional-term token probabilities to point prompts for SAM.

Problem

Existing LMM-based image quality assessment supports only part of the perception spectrum (description/QA or pixel grounding), lacking a unified task/data formulation and optimization for multi-granularity, region-aware quality understanding.

Approach

Grounding is done training-free by reading the LMM's native logits over positional terms (left/right, top/bottom), converting their softmax-weighted average into a point prompt for an off-the-shelf SAM, avoiding special seg tokens and preserving instruction-following; training is a two-stage instruction tuning of text-level tasks followed by this training-free grounding.

Key results

• On the authors' grounding benchmark, Ours (Qwen3-VL) reaches average 0.408 vs 0.264 for the best baseline Q-Instruct (adapted with text-to-point) and 0.190 for Q-Ground.
• Ours (Phi-3.5-Vision) achieves average 0.364 mIoU-style score on visual quality grounding, versus LISA(7B) 0.078 and Sa2VA(8B) 0.192.
• Human verification on 40% of the dataset: over 80% of sampled instances rated 4 or 5 on each of four dimensions.

Surprising findings

• EVF-SAM performs best when prompted with the semantic text alone and degrades with the full answer, while the training-free method with minimal spatial information outperforms it.

Limitations

• Dataset is moderate in scale (33K) and relies on InternVL-2.5-generated QA pairs.
• Grounding uses point prompts and softmax over a closed set of four positional terms, limiting spatial precision.

Experimental setup

Datasets: IQA-Spider-33K Q-Instruct DQ-495K KADIS-700K KonIQ KADID-10K Q-Bench-A1 Q-Ground-Test

Benchmarks: IQA-Spider-33K benchmark Q-Bench-A1 Q-Ground-Test KADID-10K

Metrics: GPT score accuracy mIoU SRCC PLCC

Baselines: LISA Sa2VA Q-Instruct Q-Ground EVF-SAM

Builds on: Q-Instruct DQ-495K SAM Q-Ground

Closest papers by meaning

• Grounding-IQA: Grounding Multimodal Language Model for Image Quality Assessment (ICLR 2026 · similarity 0.92)
• Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · 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)
• Refine-IQA: Multi-Stage Reinforcement Finetuning for Perceptual Image Quality Assessment (AAAI 2026 · similarity 0.88)
• Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment (ICLR 2026 · similarity 0.87)