Research map / Image generation / Editing and style transfer
Learning to Blend Photos
Wei-Chih Hung, Jianming Zhang, Xiaohui Shen, Zhe Lin, Joon-Young Lee, Ming-Hsuan Yang
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 the foreground-image branch to the quality network (ablation) | Quality Network · blending preference validation | mean accuracy: 75.23% -> 79.69% (drops by 4.46%) | Improved |
| add binary cross entropy loss to the ranking loss (ablation) | Quality Network · blending preference validation | mean accuracy: 76.17% -> 79.69% (drops 3.52%) | Improved |
| use the learned quality network instead of LPIPS | preference prediction · blending preference validation | mean accuracy: 61.78% -> 79.69% | Improved |
| increase random-search evaluation steps from 10 to 10,000 (ablation) | Random search · 20 input pairs | mean quality score: 5.69 -> 11.76 | Improved |
| replace PSO with A2C DRL search at 100 steps | search method · 20 input pairs | mean quality score: 8.91 -> 10.93 | Improved |
| replace Simulated Annealing with A2C at 100 steps | search method · 20 input pairs | mean quality score: 8.12 -> 10.93 | Improved |
| use on-policy A2C instead of Dueling-DQN (ablation) | DRL agent · 20 input pairs | mean quality score: 9.86 -> 10.93 (by 1.07) | Improved |
| use Simulated Annealing instead of Random-100 | search method · 20 input pairs | mean quality score: 8.07 -> 8.12 | No clear effect |
| use A2C search instead of PSO | search method · user study | user rating quality | Improved |
| use A2C (5 seconds) instead of Random-10k (7 more minutes) | search method · user study | user rating quality: Random-10k only a few percentage points better | No clear effect |
| use A2C search instead of the human expert | search method · user study | share of bad ratings: 14% -> 37% | Hurt |
Closest papers by meaning
- Bridging the Visual Gap: Wide-Range Image Blending (CVPR 2021 · similarity 0.87)
- Deep Image Harmonization (CVPR 2017 · similarity 0.86)
- Generative Photomontage (CVPR 2025 · similarity 0.86)
- Composition-Preserving Deep Photo Aesthetics Assessment (CVPR 2016 · similarity 0.86)
- BlendGAN: Implicitly GAN Blending for Arbitrary Stylized Face Generation (NeurIPS 2021 · similarity 0.85)
- Deep Photo Style Transfer (CVPR 2017 · similarity 0.85)
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.
