Research map / Image generation / Autoregressive and latent generation
Towards Syn-to-Real IQA: A Novel Perspective on Reshaping Synthetic Data Distributions
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 SynDR-IQA, which reshapes synthetic IQA training data via distribution-aware diverse content upsampling and density-aware redundant cluster downsampling to improve synthetic-to-real blind IQA generalization.
Problem
Blind image quality assessment models trained on synthetic distortion datasets generalize poorly to authentic and algorithmic distortions, because their learned feature spaces form discrete, clustered structures.
Approach
Attributes the clustered representations to the synthetic data distribution (limited reference content and redundant uniform distortion combinations) rather than architecture; adds diverse reference images with kNN-based pseudo-labels and downsamples dense redundant clusters, guided by a generalization bound that links diversity and redundancy heterogeneity to generalization error.
Key results
- Synthetic-to-authentic average SRCC of 0.743 (KADID-10k to LIVEC, KonIQ-10k, BID), improving over DGQA's 0.716.
- Synthetic-to-algorithmic (KADID-10k to PIPAL) average SRCC 0.5110 vs DGQA 0.4789.
- Ablation reports a 2.71% average SRCC gain over the baseline on the synthetic-to-authentic setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| reshape synthetic training data with DDCUp and DRCDown | ResNet-50 BIQA trained on KADID-10k · KADID-10k→authentic average | SRCC: +2.7% | Helped |
| use SynDR-IQA instead of DGQA | ResNet-50 BIQA trained on KADID-10k · PIPAL algorithmic average | SRCC: +3.21% | Helped |
| use SynDR-IQA instead of DGQA | ResNet-50 BIQA trained on KADID-10k · PIPAL SR and Denoising Joint | SRCC: 0.5588 -> 0.5557 (-0.31%) | Hurt |
| apply SynDR-IQA data reshaping | ResNet-50 BIQA trained on KADID-10k · KADID-10k→LIVE, CSIQ, TID2013 | SRCC: 0.9173 -> 0.9258; 0.7965 -> 0.8069; 0.7077 -> 0.7147 | Helped |
| add the full candidate dataset without selection (CD) (ablation) | baseline BIQA · KADID-10k→LIVEC and BID | SRCC: 0.6958 -> 0.6901; 0.7696 -> 0.7677 | Hurt |
| add the full candidate dataset without selection (CD) (ablation) | baseline BIQA · KADID-10k→KonIQ-10k | SRCC: 0.6810 -> 0.7105 | Helped |
| add DDCUp (CD+SEL) (ablation) | baseline BIQA · KADID-10k→LIVEC, KonIQ-10k, BID | SRCC: 0.6958 -> 0.7219; 0.6810 -> 0.7119; 0.7696 -> 0.7822 | Helped |
| add DRCDown alone (ablation) | baseline BIQA · KADID-10k→LIVEC, KonIQ-10k, BID | SRCC: 0.6958 -> 0.6962; 0.6810 -> 0.6887; 0.7696 -> 0.7793 | Helped |
| add CD and DRCDown without DDCUp selection (ablation) | baseline BIQA · KADID-10k→LIVEC, KonIQ-10k, BID | SRCC: 0.6958 -> 0.7095; 0.6810 -> 0.7107; 0.7696 -> 0.7775 | Helped |
| add CD, DDCUp selection, and DRCDown (ablation) | baseline BIQA · KADID-10k→LIVEC, KonIQ-10k, BID | SRCC: 0.6958 -> 0.7127; 0.6810 -> 0.7268; 0.7696 -> 0.7884 | Helped |
| vary DDCUp nearest-neighbor count k and distance threshold T_rf from the default (ablation) | SynDR-IQA · KADID-10k→LIVEC, KonIQ-10k, BID | SRCC: 0.713; 0.727; 0.788 (default) | Mixed |
| relax DDCUp threshold T_rf from 0.05 to 0.1 (ablation) | SynDR-IQA · KADID-10k→LIVEC, KonIQ-10k, BID | SRCC: 0.713 -> 0.704; 0.727 -> 0.713; 0.788 -> 0.779 | Hurt |
Surprising findings
- Naively adding the full candidate reference dataset improves KonIQ-10k but slightly degrades LIVEC and BID.
- Using CLIP as the full backbone gives no clear advantage over ImageNet pretraining, while using CLIP only as the DDCUp feature extractor improves results.
Limitations
- Performance gap remains on synthetic-to-algorithmic generalization (PIPAL SRCC around 0.51).
- Existing synthetic distortion types focus on traditional degradations and do not capture algorithmic distortions.
- Pseudo-labels for new samples rely on the assumption that similar content under the same distortion yields similar quality degradation.
Experimental setup
Datasets: KADID-10k LIVE CSIQ TID2013 LIVEC KonIQ-10k BID PIPAL KADIS-700k
Benchmarks: LIVEC KonIQ-10k BID PIPAL KADID-10k LIVE CSIQ TID2013
Metrics: SRCC PLCC
Baselines: RankIQA DBCNN HyperIQA MUSIQ VCRNet KGANet CLIPIQA+ Q-Align DANN UCDA RankDA StyleAM
Builds on: DGQA KADIS-700k UMAP
Closest papers by meaning
- Bridging the Synthetic-to-Authentic Gap: Distortion-Guided Unsupervised Domain Adaptation for Blind Image Quality Assessment (CVPR 2024 · similarity 0.89)
- Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption (CVPR 2025 · similarity 0.89)
- Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment (CVPR 2025 · similarity 0.88)
- Causal-IQA: Towards the Generalization of Image Quality Assessment Based on Causal Inference (ICML 2024 · similarity 0.88)
- Feature Denoising Diffusion Model for Blind Image Quality Assessment (AAAI 2025 · similarity 0.87)
- Boosting Image Quality Assessment through Efficient Transformer Adaptation with Local Feature Enhancement (CVPR 2024 · similarity 0.86)
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.