Research map / Image generation / Editing and style transfer
Photographic Image Synthesis With Cascaded Refinement Networks
Qifeng Chen, Vladlen Koltun
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 |
|---|---|---|---|
| train with VGG-19 feature matching loss instead of image-space loss (ablation) | CRN · Cityscapes | CRN preferred over baseline (%): 99.7% | Improved |
| train with VGG-19 feature matching loss instead of image-space loss (ablation) | CRN · NYU | CRN preferred over baseline (%): 91.4% | Improved |
| use cascaded refinement architecture instead of encoder-decoder (u-net) (ablation) | CRN · Cityscapes | CRN preferred over baseline (%): 78.3% | Improved |
| use cascaded refinement architecture instead of encoder-decoder (u-net) (ablation) | CRN · NYU | CRN preferred over baseline (%): 71.2% | Improved |
| use cascaded refinement architecture instead of full-resolution dilated network (ablation) | CRN · Cityscapes | CRN preferred over baseline (%): 67.7% | Improved |
| use cascaded refinement architecture instead of full-resolution dilated network (ablation) | CRN · NYU | CRN preferred over baseline (%): 65.8% | Improved |
| regression with perceptual loss instead of GAN plus semantic segmentation loss | CRN · Cityscapes | CRN preferred over baseline (%): 98.5% | Improved |
| regression with perceptual loss instead of GAN plus semantic segmentation loss | CRN · NYU | CRN preferred over baseline (%): 82.3% | Improved |
| direct regression network instead of conditional GAN of Isola et al. | CRN · Cityscapes | CRN preferred over baseline (%): 96.9% | Improved |
| direct regression network instead of conditional GAN of Isola et al. | CRN · NYU | CRN preferred over baseline (%): 77.2% | Improved |
| add diversity loss with multiple outputs (k=9) (ablation) | CRN · NYU validation layouts | diversity (qualitative) | Improved |
| compare CRN with Pix2pix at 1/8 s exposure | CRN · Cityscapes time-limited | CRN preferred over baseline (%): 73.4% | Improved |
Closest papers by meaning
- Semantic Image Synthesis With Spatially-Adaptive Normalization (CVPR 2019 · similarity 0.90)
- Semi-Parametric Image Synthesis (CVPR 2018 · similarity 0.90)
- Diverse Semantic Image Synthesis via Probability Distribution Modeling (CVPR 2021 · similarity 0.89)
- Image Synthesis via Semantic Composition (ICCV 2021 · similarity 0.89)
- PLACE: Adaptive Layout-Semantic Fusion for Semantic Image Synthesis (CVPR 2024 · similarity 0.88)
- Unlocking Pre-trained Image Backbones for Semantic Image Synthesis (CVPR 2024 · similarity 0.88)
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.
