Research map / Vision-language and multimodal / CLIP and zero-shot
QMamba: On First Exploration of Vision Mamba for Image Quality Assessment
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.
Adapts vision Mamba (VMamba/LocalMamba-style scanning) into QMamba backbones for image quality assessment and proposes StylePrompt, a lightweight mean/variance prompt tuning method for transferable IQA.
Problem
Whether Mamba (state space models) can match or surpass existing CNN and Transformer backbones for low-level perceptual tasks like image quality assessment, and how to transfer such models across IQA domains efficiently.
Approach
Builds QMamba on a VMamba-style backbone with a local window scanning scheme (LQMamba variant) so adjacent distortion-related tokens stay close, and introduces StylePrompt, which generates prompts from pooled features and uses them to predict channel-wise affine modulation parameters that adjust feature mean and variance while the backbone stays frozen.
Key results
- Task-specific IQA: LQMamba-B reports average PLCC/SRCC 0.896 across ten datasets versus 0.884 for DEIQT and 0.882 for LoDa, with QMamba-T at 4.47 GFLOPS versus 5.41 for DEIQT.
- Transferable IQA: StylePrompt tunes 3.83M parameters and reaches average 0.901 versus 0.908 for full tuning (93.79M), outperforming DEIQT (0.601) and LoDa (0.622) under the same cross-domain protocol.
- Prompt tuning ablation: StylePrompt (0.911 PLCC avg, 3.83M params) beats SSF (0.750), cross-attention prompt (0.806) and conv prompt (0.883, 28.33M params).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace cross-scanning with local window scanning (ablation) | QMamba · TID2013 | SRCC: 0.949 vs. 0.964 | Helped |
| replace cross-scanning with local window scanning (ablation) | QMamba · KADID | SRCC: 0.932 vs. 0.941 | Helped |
| increase model scale from Tiny to Small (ablation) | QMamba · task-specific IQA datasets | SRCC | Mixed |
| use dynamic style prompts rather than direct affine parameters (SSF) (ablation) | QMamba transferable IQA · cross-domain transfers | PLCC average: 0.750 -> 0.911 | Helped |
| use dynamic style prompts rather than cross-attention prompts (ablation) | QMamba transferable IQA · cross-domain transfers | PLCC average: 0.806 -> 0.911 | Helped |
| use dynamic style prompts rather than convolutional prompts (ablation) | QMamba transferable IQA · cross-domain transfers | PLCC average: 0.883 -> 0.911 | Helped |
| use six prompt components instead of one (ablation) | StylePrompt · cross-domain transfers | average: 0.8928 -> 0.9006 | Helped |
| use six prompt components instead of two (ablation) | StylePrompt · cross-domain transfers | average: 0.8936 -> 0.9006 | Helped |
| use six prompt components instead of four (ablation) | StylePrompt · cross-domain transfers | average: 0.8951 -> 0.9006 | Helped |
| use six prompt components instead of eight (ablation) | StylePrompt · cross-domain transfers | average: 0.8939 -> 0.9006 | Helped |
| use six prompt components instead of ten (ablation) | StylePrompt · cross-domain transfers | average: 0.8942 -> 0.9006 | Helped |
| use channel-only prompt shape (1,1) rather than (7,7) (ablation) | StylePrompt · cross-domain transfers | average: 0.897 -> 0.901 | Helped |
Surprising findings
- Full-model tuning and StylePrompt reach nearly equal transfer performance, while StylePrompt uses only about 4% of parameters.
- Under the reported cross-domain protocol, the frozen backbone without tuning averages 0.642, and linear probing averages 0.755, both far below tuned variants.
Limitations
- Relies on ImageNet-1K pre-trained weights, which may limit applicability to domains with different distributions.
- Evaluation focuses on specific distortion types and may not cover broader distortions; risk of overfitting to synthetic data.
- Cross-domain validity to very different domains such as medical imaging has not been validated.
Experimental setup
Datasets: LIVE CSIQ TID2013 KADID-10k LIVEC KonIQ-10k LIVEFB SPAQ AIGCIQA2023 AGIQA-3K
Benchmarks: LIVE CSIQ TID2013 KADID-10k LIVEC KonIQ-10k LIVEFB SPAQ AIGCIQA2023 AGIQA-3K
Metrics: PLCC SRCC GFLOPS
Baselines: ResNet-50 ResNet-101 ResNet-152 ViT-T ViT-S ViT-B Swin-T Swin-S Swin-B DBCNN HyperIQA TReS
Builds on: Mamba VMamba LocalMamba DEIQT SSF
Closest papers by meaning
- Demystify Mamba in Vision: A Linear Attention Perspective (NeurIPS 2024 · similarity 0.88)
- Refine-IQA: Multi-Stage Reinforcement Finetuning for Perceptual Image Quality Assessment (AAAI 2026 · similarity 0.88)
- MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025 · similarity 0.87)
- Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model (NeurIPS 2024 · similarity 0.87)
- Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.87)
- Vision Mamba Mender (NeurIPS 2024 · 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.