Research map / Image generation / Autoregressive and latent generation
Learning Conditional Knowledge Distillation for Degraded-Reference Image Quality Assessment
Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, Jiebo Luo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add conditional knowledge distillation (CKD) loss (ablation) | CKDN · PIPAL subset (50 train / 39 val IR algorithms) | SRCC: 0.7127 -> 0.7549 | Improved |
| add relative score regression pre-training (ablation) | CKDN · PIPAL subset (50 train / 39 val IR algorithms) | SRCC: 0.7549 -> 0.7669 | Improved |
| use unshared DTE and QSE embeddings instead of shared (ablation) | CKDN · PIPAL subset (50 train / 39 val IR algorithms) | SRCC: 0.7311 -> 0.7669 | Improved |
| keep full 72x72 feature resolution for CSP input instead of 36x36 (ablation) | CKDN · PIPAL subset (50 train / 39 val IR algorithms) | SRCC: 0.7436 -> 0.7669 | Improved |
| keep full 72x72 feature resolution instead of 18x18 (ablation) | CKDN · PIPAL subset (50 train / 39 val IR algorithms) | SRCC: 0.7234 -> 0.7669 | Improved |
| use residual blocks instead of VGG blocks (ablation) | CKDN · PIPAL subset (50 train / 39 val IR algorithms) | SRCC: 0.7006 -> 0.7127 | No clear effect |
| use conditional KD instead of vanilla knowledge distillation (ablation) | CKDN · PIPAL subset (50 train / 39 val IR algorithms) | SRCC: 0.7338 -> 0.7669 | Improved |
| use conditional KD instead of data augmentation (ablation) | CKDN · PIPAL subset (50 train / 39 val IR algorithms) | SRCC: 0.7469 -> 0.7669 | Improved |
| use conditional KD instead of hallucinated reference (ablation) | CKDN · PIPAL subset (50 train / 39 val IR algorithms) | SRCC: 0.7348 -> 0.7669 | Improved |
| pre-train with relative score regression instead of absolute score regression (10 epochs) (ablation) | CKDN · PIPAL subset (50 train / 39 val IR algorithms) | SRCC: 0.7536 -> 0.7669 | Improved |
| pre-train with relative score regression instead of ImageNet pre-training only (ablation) | CKDN · PIPAL subset (50 train / 39 val IR algorithms) | SRCC: 0.7549 -> 0.7669 | Improved |
| add CKDN as loss term to ESRGAN instead of VGG perceptual loss | ESRGAN · DIV8K SR (ESRGAN) | LPIPS: 0.5039 -> 0.4957 | Improved |
Closest papers by meaning
- Comparison of Full-Reference Image Quality Models for Optimization of Image Processing Systems (IJCV 2021 · similarity 0.90)
- UARE: A Unified Vision-Language Model for Image Quality Assessment, Restoration, and Enhancement (CVPR 2026 · similarity 0.88)
- Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment (CVPR 2025 · similarity 0.88)
- Adaptive Feature Selection for No-Reference Image Quality Assessment by Mitigating Semantic Noise Sensitivity (ICML 2024 · similarity 0.88)
- Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption (CVPR 2025 · similarity 0.87)
- Causal-IQA: Towards the Generalization of Image Quality Assessment Based on Causal Inference (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.
