Image Restoration - Approaches and Models¶

Overview of image restoration approaches: from classical to diffusion-based. Key insight: diffusion models bring generalization across degradation types but at higher compute cost.

Classical / CNN-Based¶

SwinIR (ICCV 2021)¶

• Swin Transformer blocks for image restoration
• Strong baseline for denoise, super-res, JPEG artifact removal
• Lightweight (~12M params), fast inference
• Limitation: task-specific training, one model per degradation

NAFNet (ECCV 2022)¶

• "Nonlinear Activation Free Network" - removes GELU/Softmax
• SOTA on SIDD (40.30 dB PSNR) and GoPro deblurring
• Very efficient: ~67M params, simple architecture
• Uses SimpleGate + Simplified Channel Attention

Restormer (CVPR 2022)¶

• Multi-scale Transformer for high-res restoration
• Transposed attention: key/value along channel dim (not spatial)
• Strong on real noise removal, motion deblur, defocus deblur

Diffusion-Based¶

[[RealRestorer]] (March 2026)¶

• 9 degradation types on [[Step1X-Edit]] backbone
• Prompt-driven: specify degradation type in text

• 1 open-source on RealIR-Bench (FS=0.146), close to GPT-Image-1.5¶

• ~34 GB VRAM, 28 steps
• Weights: non-commercial academic only

Palette (Google, 2022)¶

• First diffusion model for image-to-image restoration
• Concatenates degraded image with noise as conditioning
• Showed diffusion can match/beat task-specific models

IR-SDE (NeurIPS 2023)¶

• Treats restoration as SDE reverse process
• Mean-reverting SDE: starts from degraded image, not pure noise
• Better than starting from noise for restoration tasks

Degradation Types¶

SANA-Denoiser Approach¶

Our approach: repurpose [[SANA]] 1.6B DiT as restoration model via [[Paired Training for Restoration]]: - Channel concat conditioning (degraded latent + noise) - [[DC-AE]] 32x compression keeps token count low - Linear attention O(N) enables high-res processing - [[Temporal Tiling]] for context-aware tile processing at 4K+

Advantages over RealRestorer: - 10x fewer params (1.6B vs ~15B Step1X-Edit backbone) - Linear attention vs quadratic (much faster at high-res) - 32x VAE compression vs 8x (4x fewer tokens)

Standard Benchmarks¶