Research map / Vision-language and multimodal / Multimodal LLMs

Revisiting MLLM Based Image Quality Assessment: Errors and Remedy

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs · 1 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 Q-Scorer, an MLLM-based image quality assessment method that replaces level tokens with IQA-specific score tokens plus a lightweight MLP regression head, with LoRA fine-tuning on mPLUG-Owl2.

Problem

MLLM-based IQA methods map discrete level tokens (e.g., "good", "poor") to continuous quality scores, which causes conversion errors and semantic confusion from pre-trained token meanings and from IQA fine-tuning overwriting token semantics.

Approach

Introduce a set of IQA-specific score tokens (score1..score5, one per score interval) into the LLM vocabulary, and regress a continuous score from the selected score token's embedding with an MLP, trained jointly with cross-entropy and MSE losses.

Key results

• Trained only on KonIQ, Ours (5) reaches PLCC/SRCC 0.959/0.948 on KonIQ, improving over fully fine-tuned Q-Align (0.941/0.940) and LoRA Q-Align (0.932/0.938).
• Cross-dataset generalization: Ours (1) reaches 0.903/0.888 on LIVE-Wild and 0.900/0.899 on SPAQ, outperforming DeQA-Score on most unseen datasets except KADID.
• Multi-dataset training with fidelity loss reaches 0.963/0.961 on KADID-inclusive setting (Ours F-loss, KonIQ+SPAQ+KADID) versus DeQA-Score 0.955/0.953 on KADID.

Surprising findings

• A single shared score token (Ours(1)) outperforms the five interval-specific score tokens in single-dataset training, while it significantly degrades in multi-dataset settings.
• Replacing the MLP regressor with a probability-weighted sum over score tokens causes a significant performance drop, and the paper reports that Q-Align-style level tokens suffer post-tuning degradation on text-to-image alignment assessment.

Limitations

• Most IQA datasets reflect annotators' personal preferences, so models may not capture general human perception.
• Multi-dataset training with a unified score token set and single MLP head can cause interference, hurting KonIQ performance without fidelity loss.
• Score tokens are sometimes assigned cumulative probability well below 1, indicating suboptimal alignment of new tokens.

Experimental setup

Datasets: KonIQ SPAQ KADID

Benchmarks: KonIQ SPAQ KADID LIVE-Wild AGIQA-3K CSIQ

Metrics: PLCC SRCC

Baselines: Q-Align DeQA-Score Q-Align (LoRA) Compare2Score VideoScore NIQE BRISQUE NIMA HyperIQA DBCNN MUSIQ CLIP-IQA+

Builds on: Q-Align DeQA-Score mPLUG-Owl2 VideoScore Compare2Score

Closest papers by meaning

• Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution (CVPR 2025 · similarity 0.94)
• Grounding-IQA: Grounding Multimodal Language Model for Image Quality Assessment (ICLR 2026 · similarity 0.91)
• IQA-Spider: Unifying Multi-Granularity Image Quality Assessment with Reasoning, Grounding and Referring (ICML 2026 · similarity 0.89)
• Q-Bench: A Benchmark for General-Purpose Foundation Models on Low-level Vision (ICLR 2024 · similarity 0.89)
• Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.89)
• Self-Evolving Vision-Language Models for Image Quality Assessment via Voting and Ranking (ICLR 2026 · similarity 0.87)