Reproduce “Contribution-based Low-Rank Adaptation with Pre-training Model for Real Image Restoration” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Donwon Park, Hayeon Kim, Se Young Chun

A method for real image restoration integrates Random Order Degradation (PROD) for pre-training and Contribution-based Low-Rank Adaptation (CoLoRA) for fine-tuning. This approach achieves comparable restoration quality to full fine-tuning using approximately 7% of the parameters, performing well across diverse degradations and with limited training data.

Key insights

  • Pre-training with dynamically ordered and combined degradations (PROD) helps models generalize better to the diverse and complex nature of real-world image degradations.
  • Layer-specific contribution analysis allows for more efficient parameter allocation in low-rank adaptation, leading to performance comparable to full fine-tuning with significantly fewer parameters.
  • Encoder and decoder layers often have higher contributions to image restoration tasks, making them prime targets for selective parameter allocation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.