Research map / Image generation / Autoregressive and latent generation
Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption
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.
Builds DiffIQA (~180k diffusion-enhanced test images with human worse/similar/better labels vs. reference) and SRIQA-Bench, and proposes A-FINE, a full-reference IQA model that adaptively combines a DISTS-like fidelity term and a naturalness term so it works when the reference is imperfect.
Problem
Full-reference IQA models assume the reference image has perfect quality, which fails when references are degraded or when enhancement methods produce outputs better than the reference.
Approach
Model quality as an adaptive linear combination of a fidelity term (DISTS-like, on CLIP ViT features) and a naturalness term of the test image, with the weight driven by the difference in naturalness between reference and test; it reduces to standard FR-IQA when the reference is much more natural than the test image.
Key results
- A-FINE reaches 81.8% average accuracy across the combined test sets, versus 78.8% for fine-tuned TOPIQ and 74.2% for fine-tuned CKDN.
- On SRIQA-Bench, A-FINE achieves 82.4% overall accuracy, above the best fine-tuned baseline DISTS-FT (71.4%) and TOPIQ-FT (77.7%).
- A-FINE gets 79.9% on DiffIQA, with 82.3% on Ref>Test and 79.9% on the all-pairs average in the table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use A-FINE instead of TOPIQ | FR-IQA evaluation on TID2013 · TID2013 test | accuracy (%): 90.4 -> 88.1 | Hurt |
| use A-FINE instead of TOPIQ | FR-IQA evaluation on KADID-10K · KADID-10K test | accuracy (%): 94.3 -> 88.3 | Hurt |
| use A-FINE instead of TOPIQ | FR-IQA evaluation on PIPAL · PIPAL test | accuracy (%): 80.5 -> 81.0 | Helped |
| use A-FINE instead of TOPIQ-FT | FR-IQA evaluation on DiffIQA · DiffIQA average | accuracy (%): 77.0 -> 79.9 | Helped |
| use A-FINE instead of TOPIQ-FT | FR-IQA evaluation on SRIQA-Bench · SRIQA-Bench generation-based | accuracy (%): 73.0 -> 78.9 | Helped |
| replace CLIP ViT-B/32 with VGG16 (ablation) | A-FINE · Standard IQA average | accuracy (%): 83.6 -> 77.6 | Hurt |
| replace CLIP ViT-B/32 with ImageNet-trained ResNet50 (ablation) | A-FINE · Standard IQA average | accuracy (%): 83.6 -> 74.8 | Hurt |
| replace CLIP ViT-B/32 with CLIP-trained ResNet50 (ablation) | A-FINE · Standard IQA average | accuracy (%): 83.6 -> 76.1 | Hurt |
| replace CLIP ViT-B/32 with ImageNet-trained ViT-B/32 (ablation) | A-FINE · Standard IQA average | accuracy (%): 83.6 -> 81.0 | Hurt |
| train on standard IQA datasets only, without DiffIQA (ablation) | A-FINE · DiffIQA average | accuracy (%): 79.9 -> 65.6 | Hurt |
| train on DiffIQA only, without standard IQA datasets (ablation) | A-FINE · Standard IQA average | accuracy (%): 83.6 -> 70.6 | Hurt |
| replace combined training data with standard IQA datasets only (ablation) | A-FINE · SRIQA-Bench all | accuracy (%): 82.4 -> 78.7 | Hurt |
Surprising findings
- Fine-tuning on DiffIQA typically lowers performance of standard FR-IQA models on standard IQA datasets, indicating a trade-off between distorted and enhanced image evaluation.
- Standard FR-IQA models such as PSNR and DISTS drop to near-chance or below on DiffIQA pairs where the test image is better than the reference (e.g., PSNR 18.2% on DiffIQA average).
Limitations
- When the reference image quality is poor, the fidelity term can introduce significant bias in quality prediction.
- A observed performance trade-off between evaluating distorted versus enhanced images.
- A-FINE is an early effort; the optimal functional form of asymmetric distance measures remains open.
Experimental setup
Datasets: TID2013 KADID-10K PIPAL DiffIQA DF2K DIV8K FFHQ LSDIR SRIQA-Bench
Benchmarks: TID2013 KADID-10K PIPAL DiffIQA SRIQA-Bench
Metrics: accuracy (%)
Baselines: PSNR SSIM MS-SSIM FSIM VSI LPIPS DISTS AHIQ TOPIQ VIF PCQI SFSN
Builds on: DISTS PASD UNIQUE CKDN Thurstone's Case V model
Closest papers by meaning
- Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.91)
- Augmenting Perceptual Super-Resolution via Image Quality Predictors (CVPR 2025 · similarity 0.89)
- Towards Syn-to-Real IQA: A Novel Perspective on Reshaping Synthetic Data Distributions (NeurIPS 2025 · similarity 0.89)
- Causal-IQA: Towards the Generalization of Image Quality Assessment Based on Causal Inference (ICML 2024 · similarity 0.88)
- UARE: A Unified Vision-Language Model for Image Quality Assessment, Restoration, and Enhancement (CVPR 2026 · similarity 0.88)
- Adaptive Feature Selection for No-Reference Image Quality Assessment by Mitigating Semantic Noise Sensitivity (ICML 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.