Research map / Image generation / Image restoration
Feature Denoising Diffusion Model for Blind 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.
Proposes PFD-IQA, a blind image quality assessment model that uses a text-conditioned lightweight diffusion model (DDIM, ~5 steps) to denoise ViT features, guided by perceptual text prompts built from distortion-type and quality-level priors.
Problem
Pre-trained high-level features used for transfer learning in blind image quality assessment contain quality-irrelevant noise, causing models to focus on semantics rather than distortions.
Approach
Treat student features as noisy versions of teacher pseudo-features and denoise them with a text-conditioned diffusion model along predefined trajectories, using an adaptive noise-level alignment to set the starting timestep and perceptual prompts (distortion and quality-level text embeddings weighted by predicted similarity) as conditions.
Key results
- PLCC of 0.935 on KADID (up 3.0% over 0.905 of DACNN) and 0.922 on LIVEC (up 2.8% over 0.894 of DEIQT).
- Achieves best PLCC/SRCC on all eight datasets in the main comparison table, e.g., 0.945 PLCC on KonIQ and 0.667 PLCC on LIVEFB.
- Ablation: PDA and PDR together raise the average over the five-dataset ablation from 0.911 (baseline) to 0.946; 5 sampling steps performs best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the PDA module (ablation) | PFD-IQA · LIVEC | PLCC: 0.881 -> 0.911 | Helped |
| add the PDR module (ablation) | PFD-IQA · LIVEC | PLCC: 0.881 -> 0.916 | Helped |
| combine PDA and PDR (ablation) | PFD-IQA · LIVEC | PLCC: 0.881 -> 0.922 | Helped |
| add adaptive noise alignment (ANA) (ablation) | PFD-IQA PDR · LIVEC | PLCC: 0.881 -> 0.914 | Helped |
| add perceptual prompt aggregation (PPA) (ablation) | PFD-IQA PDR · LIVEC | PLCC: 0.881 -> 0.918 | Helped |
| combine ANA and PPA (ablation) | PFD-IQA PDR · LIVEC | PLCC: 0.881 -> 0.922 | Helped |
| use 5 rather than 1 sampling step (ablation) | PFD-IQA · LIVE | PLCC: 0.982 -> 0.985 | Helped |
| use 5 rather than 3 sampling steps (ablation) | PFD-IQA · LIVE | PLCC: 0.984 -> 0.985 | Helped |
| replace DEIQT with PFD-IQA | BIQA model · KADID | PLCC: 0.905 -> 0.935 | Helped |
| replace DEIQT with PFD-IQA | BIQA model · LIVEC | PLCC: 0.894 -> 0.922 | Helped |
Surprising findings
- Single-step denoising already substantially outperforms the baseline, and performance saturates at around 5 sampling iterations.
Limitations
- Relies on a pre-trained teacher model to generate pseudo-features and pseudo-labels, so performance depends on teacher quality.
- Cross-dataset generalization gains are reported only on a limited set of dataset pairs.
Experimental setup
Datasets: LIVE CSIQ TID2013 KADID LIVEC KonIQ LIVEFB SPAQ
Benchmarks: LIVE CSIQ TID2013 KADID LIVEC KonIQ LIVEFB SPAQ
Metrics: PLCC SRCC
Baselines: DIIVINE BRISQUE ILNIQE BIECON MEON WaDIQaM DBCNN TIQA MetaIQA P2P-BM HyperIQA TReS
Builds on: DDPM DDIM DEIQT CLIP-based text prompts
Closest papers by meaning
- Integrating Global Context Contrast and Local Sensitivity for Blind Image Quality Assessment (ICML 2024 · similarity 0.88)
- Towards Syn-to-Real IQA: A Novel Perspective on Reshaping Synthetic Data Distributions (NeurIPS 2025 · similarity 0.87)
- Semi-Supervised Blind Image Quality Assessment through Knowledge Distillation and Incremental Learning (AAAI 2024 · similarity 0.87)
- Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment (CVPR 2025 · similarity 0.86)
- LGDM: Latent Guidance in Diffusion Models for Perceptual Evaluations (ICML 2025 · similarity 0.86)
- Adaptive Feature Selection for No-Reference Image Quality Assessment by Mitigating Semantic Noise Sensitivity (ICML 2024 · similarity 0.85)
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.