Reproduce “Dual-domain Adaptation Networks for Realistic Image Super-resolution” on OpenResearch

About this paper

Chaowei Fang, Bolin Fu, De Cheng, Lechao Cheng, Guanbin Li

Dual-domain Adaptation Networks (DAN) for realistic image super-resolution efficiently adapt pre-trained models by combining selective spatial parameter fine-tuning with Low-Rank Adaptation (LoRA) and a dedicated frequency-domain processing branch. This method achieves state-of-the-art performance on realistic SR benchmarks with significantly fewer trainable parameters and faster training times.

Key insights

• Efficient adaptation of pre-trained models, through selective parameter updates and low-rank adaptation, is crucial for bridging the gap between synthetic and realistic SR tasks with limited data.
• Explicitly processing high-frequency components in the frequency domain, in addition to spatial-domain processing, leads to superior recovery of fine details and sharper image reconstruction.
• The strategic combination of freezing shallow units and adapting deeper units within transformer groups, augmented by LoRA, strikes an effective balance between preserving learned features and domain-specific adaptation.

Summary by alphaXiv.