Research map / Image generation / Super-resolution
Adaptive Dropout: Unleashing Dropout across Layers for Generalizable Image Super-Resolution
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 Adaptive Dropout, a regularizer for blind super-resolution that couples dropped and un-dropped features with a learnable weight and anneals it layer-wise, so dropout can be applied to intermediate layers.
Problem
Blind super-resolution models overfit to training degradations; standard dropout harms performance when applied to intermediate layers due to training-testing inconsistency and differing generalization needs across layers.
Approach
Re-designs dropout as an adaptive coupling w*x + (1-w)*dropout(x) that reduces variance shift by (1-w)^2, plus a layer-wise annealing training strategy (implicit learned w via channel-attention-like module, or explicit block-by-block setting of w to 1 from shallow to deep).
Key results
- Explicit Adaptive Dropout improves SRResNet on Set5 clean from 25.67 (standard Dropout) to 26.11 dB at x4 and SwinIR on Urban100 clean from 22.27 to 22.41 dB.
- On real-world datasets, RRDB with Explicit Adaptive Dropout reaches 25.39 dB on RealSR mild versus 24.76 dB for Simple-Align.
- Adaptive dropout applied to RealESRGAN raises PSNR from 23.19 to 23.64 dB and lowers LPIPS from 0.31 to 0.24.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add adaptive dropout rather than Simple-Align | RRDB · RealSR | PSNR: 24.76 -> 25.39 | Helped |
| Add adaptive dropout rather than Simple-Align | RRDB · DRealSR difficult | PSNR: 17.14 -> 17.62 | Helped |
| Add adaptive dropout rather than Simple-Align | RRDB · Urban100 | PSNR: 21.05 -> 21.00 | Hurt |
| Add adaptive dropout format (ablation) | SRResNet blind SR · blind SR ablation | PSNR: 24.33 -> 25.89 | Helped |
| Add adaptive training strategy (ablation) | SRResNet with adaptive dropout · blind SR ablation | PSNR: 25.89 -> 26.07 | Helped |
| Use standard dropout instead of adaptive dropout (ablation) | SRResNet with adaptive training strategy · blind SR ablation | PSNR: 26.07 -> 24.89 | Hurt |
| Set implicit adaptive dropout w to 0.7 (ablation) | SRResNet · blind SR ablation | PSNR: 25.83 -> 26.07 | Helped |
| Increase implicit adaptive dropout w from 0.7 to 0.9 (ablation) | SRResNet · blind SR ablation | PSNR: 26.07 -> 25.86 | Hurt |
Surprising findings
- Directly applying standard dropout to intermediate layers degrades performance, and stronger regularization worsens channel imbalance at intermediate layers.
Limitations
- Applying multiple adaptive dropout layers within the same CNN block does not improve performance on unseen degradations.
- Hyperparameter w must be tuned (best around 0.7-0.8 for implicit variant); p fixed at 0.5.
Experimental setup
Datasets: DIV2K Set5 Set14 BSD100 Manga109 Urban100 NTIRE 2018 RealSR DRealSR Rain100L Rain100H ITS
Benchmarks: Set5 Set14 BSD100 Manga109 Urban100 NTIRE 2018 RealSR DRealSR Rain100H OTS
Metrics: PSNR SSIM LPIPS MANIQA SRGA
Baselines: Dropout Simple-Align No regularization
Builds on: Dropout Simple-Align Real-ESRGAN
Closest papers by meaning
- Reflash Dropout in Image Super-Resolution (CVPR 2022 · similarity 0.89)
- Navigating Beyond Dropout: An Intriguing Solution towards Generalizable Image Super Resolution (CVPR 2024 · similarity 0.89)
- AutoLUT: LUT-Based Image Super-Resolution with Automatic Sampling and Adaptive Residual Learning (CVPR 2025 · similarity 0.85)
- DREAM: Diffusion Rectification and Estimation-Adaptive Models (CVPR 2024 · similarity 0.82)
- Everything at Every Scale: Scale-Invariant Diffusion with Continuous Super-Resolution (NeurIPS 2026 · similarity 0.82)
- Towards Realistic Data Generation for Real-World Super-Resolution (ICLR 2025 · similarity 0.81)
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.