Research map / Image generation / Image restoration
Analyzing the Training Dynamics of Image Restoration Transformers: A Revisit to Layer Normalization
MinKyu Lee, Sangeek Hyun, Woojin Jun, Hyunjun Kim, Jiwoo Chung, Jae-Pil Heo
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 |
|---|---|---|---|
| replace LayerNorm with i-LN | HAT_1 · x4 SR Urban100 | PSNR: 26.55 -> 27.17 | Improved |
| replace LayerNorm with i-LN | HAT_1 · Rain100L deraining | PSNR: 34.35 -> 36.20 | Improved |
| replace LayerNorm with i-LN | HAT_1 · Urban100 denoising sigma=15 | PSNR: 35.489 -> 35.558 | Improved |
| replace LayerNorm with i-LN | DRCT_1 · x2 SR Set5 | PSNR: 38.19 -> 38.23 | No clear effect |
| add spatial holisticness (ablation) | i-LN · x4 SR Urban100 (HAT_2) | PSNR: 26.8779 -> 27.5335 | Improved |
| add rescaling to spatial holistic norm (ablation) | i-LN · x4 SR Urban100 (HAT_2) | PSNR: 27.5335 -> 27.5849 | Improved |
| use InstanceNorm instead of i-LN (ablation) | HAT_1 · x4 SR Urban100 | PSNR: 27.17 -> 27.02 | Hurt |
| use per-token variants (LayerScale, RMSNorm) instead of LN (ablation) | HAT_1 · x4 SR Urban100 | PSNR: 26.55 -> 26.75 (LayerScale) | Improved |
| use ReZero instead of LN (ablation) | HAT_1 · x4 SR Urban100 | PSNR: 26.55 -> 26.41 | Hurt |
| remove all normalization (ablation) | HAT_1 · x4 SR | PSNR | Hurt |
| replace LN with i-LN under int4 weight quantization | HAT_2 · x4 SR Urban100 | PSNR: 25.0242 -> 26.8292 | Improved |
| replace LN with i-LN under fp16 inference | HAT_2 · x4 SR Urban100 | PSNR: 7.4640 -> 27.5849 | Improved |
Closest papers by meaning
- Alleviating Distortion in Image Generation via Multi-Resolution Diffusion Models and Time-Dependent Layer Normalization (NeurIPS 2024 · similarity 0.87)
- Impact of Layer Norm on Memorization and Generalization in Transformers (NeurIPS 2025 · similarity 0.87)
- Neural Collapse is Globally Optimal in Deep Regularized ResNets and Transformers (NeurIPS 2025 · similarity 0.86)
- Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration (NeurIPS 2026 · similarity 0.86)
- Debiased All-in-one Image Restoration with Task Uncertainty Regularization (AAAI 2025 · similarity 0.86)
- SLAB: Efficient Transformers with Simplified Linear Attention and Progressive Re-parameterized Batch Normalization (ICML 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.
