VisualQuality-Bench

This repository records common image processing tasks with their synthetic data construction procedures.

Image Processing Tasks

Image Restoration Tasks

• Super-Resolution
  • Synthetic data construction
    • For non-generative methods: Bicubic downsampling (x2, x4)
    • For generative methods: RealESRGAN degradation [RealESRGAN Paper]
      
  • SOTA methods
    • SwinIR, HAT, DiffBIR, SeeSR, OSEDiff, HyperIR
• Denoising
  • Synthetic data construction
    • Camera perception: [Unprocessing pipeline]
    • CBDNet degradation: [CBDNet] (the most common)
      
  • SOTA methods
    • CBDNet, SwinIR, Restormer, X-Restormer,
• Deblurring
  • Synthetic data construction
    • [Levin Dataset Kernels]
    • Gaussian kernel + motion kernel + noise
  • SOTA methods
    • MAXIM, LaKDNet, EVSSM,
• Deraining
  • Synthetic data construction
    • Rain streak, Raindrop, Rain and mist [MPID Dataset]
  • SOTA methods
    • DiffPlugin, Restormer, MPRNet,
• Desnowing
  • Synthetic data construction
    • Manual mask + random overlay + brightness change [Snow100K]
  • SOTA methods
    • DesnowNet,
• Dehazing
  • Synthetic data construction
    • Atmospheric Scattering Model [RESIDE]
      
  • SOTA methods
    • DehazeFormer (2023), RIDCP (2023),
• Natural Image Inpainting
  • Synthetic data construction
    • Free-form irregular masks and rectangular masks [SN-PatchGAN]
  • SOTA methods
    • SN-PatchGAN (2019),
• Reflection Removal
  • Synthetic data construction
    • [Zhang et al.]
      
  • SOTA methods
    • CEILNet (2017), Zhang et al. (2018),
• Old Photo Restoration
  • Synthetic data construction
    • Unstructured Degradation (noise) + Structured Degradation (Scratches, holes/tears) [Pipeline]
  • SOTA methods
    • Wan et al. (2020),
• Face Inpainting
  • Synthetic data construction
    • Free-form irregular masks and rectangular masks
  • SOTA methods
    • RePaint (2022), CodeFormer (2022),
• Face Super-Resolution
  • Synthetic data construction
    • GFPGAN degradation [Pipeline]
  • SOTA methods
    • GFPGAN (2021), CodeFormer (2022),

Image Enhancement Tasks

• Low-Light Enhancement
  • Synthetic data construction
    • Reversed ISP + Exposure Degradation + Noise Degradation + White Balance & CCM + Tone Mapping + Forward ISP [SynLLIE]
  • SOTA methods
    • SNR-LLIE (2022), Retinexformer (2023),
• Color Enhancement
  • Synthetic data construction
    • RAW to ISP, Changing EV [Afifi et al.]
  • SOTA methods
    • Afifi et al. (2021),
• Underwater Image Enhancement
  • Synthetic data construction
    • [Jaffe-McGlamery / Sea-thru]
    • WaterGAN (complicated)
  • SOTA methods
    • FUnIE-GAN (2020),
• Multi-Exposure Fusion
  • Synthetic data construction
    • Inverse Gamma + Changing EV + CRF and noise
  • SOTA methods
    • U2Fusion (2020)

Tone Mapping Tasks

• HDR Tone Mapping
  • Synthetic data construction
    • HDR datasets
  • SOTA methods
    • TMO algorithms

Image Generation Tasks

• Style Transfer (I2I)
  • Content Image and Style Image
  • CycleGAN (2017)
• Text-to-Image Synthesis
  • GPT generation with attribution combinations
  • SD3.5 (2024),
• Image Editing
  • Send images into a VLM to generate instructions
  • SD3.5 (2024),
• Image Outpainting
  • Natural Image
  • QueryOTR (2022)

Compression Tasks

• Image Compression
  • JPEG, WebP, HEIC, Learned algorithms