Research map / Image generation / Image restoration
Scaling Laws For Deep Learning Based Image Reconstruction
Tobit Klug, Reinhard Heckel
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 |
|---|---|---|---|
| Scale SwinIR to 128.9M parameters and train on 100k images | SwinIR · CBSD68, Kodak24, McMaster, Urban100 | PSNR: 0.05 to 0.22dB | Improved |
| Increase denoising training data beyond a few thousand examples | U-Net and SwinIR · ImageNet denoising test | PSNR scaling coefficient: 0.0048 to 0.0019; 0.0050 to 0.0017 | Mixed |
| Increase MRI training data beyond about 2.5k examples | U-Net · fastMRI multi-coil brain (4x acceleration) | SSIM scaling coefficient: 0.0072 to 0.0026 | Mixed |
| Increase super-resolution training data beyond roughly 30k examples | U-Net · ImageNet super-resolution (factor 2) | PSNR scaling coefficient: 0.0075 to 0.0029 | Mixed |
| Use a SwinIR architecture instead of a U-Net | ImageNet denoising · ImageNet denoising test | PSNR | Improved |
| Use a smaller noise variance, σz=15 instead of σz=25 (ablation) | U-Net denoiser · ImageNet denoising | PSNR: about 1.3dB | Improved |
| Fix the noise per training image instead of re-sampling it each epoch (ablation) | U-Net denoiser · ImageNet denoising | PSNR: 0.3dB | Hurt |
| Add DIV2K, Flickr2K, BSD500, and WED images to the training data (ablation) | U-Net denoiser · ImageNet denoising test | PSNR: 32.252dB -> 32.239dB | Hurt |
| Use 128×128 patches instead of 256×256 patches (ablation) | U-Net denoiser · ImageNet denoising test | PSNR | Hurt |
| Increase MRI optimizer choice from RMSprop to Adam (ablation) | U-Net MRI reconstructor · fastMRI multi-coil brain (4x acceleration) | SSIM | No clear effect |
| Increase batch size beyond 10 (ablation) | U-Net denoiser · ImageNet denoising test | PSNR | No clear effect |
| Increase network size for the large-training-set MRI settings (ablation) | U-Net MRI reconstructor · fastMRI multi-coil brain (4x acceleration) | SSIM | No clear effect |
Closest papers by meaning
- Scaling Laws in Linear Regression: Compute, Parameters, and Data (NeurIPS 2024 · similarity 0.87)
- Deep Linear Network Training Dynamics from Random Initialization: Data, Width, Depth, and Hyperparameter Transfer (ICML 2025 · similarity 0.87)
- A Dynamical Model of Neural Scaling Laws (ICML 2024 · similarity 0.87)
- Depthwise Hyperparameter Transfer in Residual Networks: Dynamics and Scaling Limit (ICLR 2024 · similarity 0.87)
- Navigating Scaling Laws: Compute Optimality in Adaptive Model Training (ICML 2024 · similarity 0.86)
- Scaling Laws for Sparsely-Connected Foundation Models (ICLR 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.
