Research map / Image generation / Super-resolution
Inpainting at Modern Camera Resolution by Guided PatchMatch with Auto-Curation
Lingzhi Zhang, Connelly Barnes, Kevin Wampler, Sohrab Amirghodsi, Eli Shechtman, Zhe Lin, Jianbo Shi
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 |
|---|---|---|---|
| add pretraining on real-vs-fake classification (ablation) | curation network · Human paired preference data | accuracy: 53.9% -> 56.4% | Improved |
| add JPEG, rotation and noise augmentations (ablation) | curation network · Human paired preference data | accuracy: 53.8% -> 56.4% | Improved |
| input the hole mask (ablation) | curation network · Human paired preference data | accuracy: 52.6% -> 56.4% | Improved |
| use shared-backbone pairwise architecture instead of early fusion (ablation) | curation network · Human paired preference data | accuracy: 51.4% -> 56.4% | Improved |
| fine-tune the backbone instead of freezing it (ablation) | curation network · Human paired preference data | accuracy: 43.7% -> 56.4% | Improved |
| use curation network instead of random choice among 8 guided PatchMatch results (ablation) | SuperCAF · 4K+ inpainting benchmark, 100 user-study photos | FID: 19.7704 -> 18.7414 | Improved |
| use SuperCAF instead of Photoshop Content-Aware Fill | SuperCAF · 4K+ inpainting benchmark, 100 user-study photos | FID: 23.0068 -> 18.7414 | Improved |
| replace Real-ESRGAN upsampling with our guided PatchMatch framework on LaMa | LaMa · 4K+ inpainting benchmark | FID: 19.2022 -> 18.7414 | Improved |
| replace Real-ESRGAN upsampling with our framework on CoModGAN | CoModGAN · 4K+ inpainting benchmark | FID: 24.8189 -> 19.9224 | Improved |
| run deep inpainting on full image instead of auto-crop (ablation) | CoModGAN · 4K+ inpainting benchmark | FID: 25.05 -> 24.8 | No clear effect |
| use Real-ESRGAN instead of bicubic upsampling (ablation) | LaMa · 4K+ inpainting benchmark | FID: 23.66 -> 19.20 | Improved |
| use Real-ESRGAN instead of reference-based TTSR upsampling (ablation) | LaMa · 4K+ inpainting benchmark (650-image subset) | FID: 25.93 -> 25.03 | Improved |
Closest papers by meaning
- High-Resolution Image Inpainting with Iterative Confidence Feedback and Guided Upsampling (ECCV 2020 · similarity 0.90)
- Ultra High-Resolution Image Inpainting with Patch-Based Content Consistency Adapter (ICCV 2025 · similarity 0.89)
- High-Resolution Image Inpainting Using Multi-Scale Neural Patch Synthesis (CVPR 2017 · similarity 0.89)
- Reconstructing the Image Stitching Pipeline: Integrating Fusion and Rectangling into a Unified Inpainting Model (NeurIPS 2024 · similarity 0.88)
- Towards Enhanced Image Inpainting: Mitigating Unwanted Object Insertion and Preserving Color Consistency (CVPR 2025 · similarity 0.86)
- IMFine: 3D Inpainting via Geometry-guided Multi-view Refinement (CVPR 2025 · 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.
