Research map / Image generation / Editing and style transfer
Unsupervised Image-to-Image Translation with Stacked Cycle-Consistent Adversarial Networks
Minjun Li, Haozhi Huang, Lin Ma, Wei Liu, Tong Zhang, Yugang Jiang
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 Stage-2 refinement (128-256) on top of Stage-1 128 (ablation) | SCAN · Cityscapes Labels->Photo | pixel accuracy: 0.46 -> 0.64 | Improved |
| add Stage-2 refinement (128-256) on top of Stage-1 128 (ablation) | SCAN · Cityscapes Photo->Labels | pixel accuracy: 0.72 -> 0.71 | Hurt |
| train at low resolution 128 instead of 256 for Stage-1 (ablation) | SCAN Stage-1 · Cityscapes Photo->Labels | pixel accuracy: 0.63 -> 0.71 | Improved |
| double parameters with Stage-2 at 256-256 versus Stage-1 256 (ablation) | SCAN · Cityscapes Labels->Photo | pixel accuracy: 0.57 -> 0.52 | No clear effect |
| train both stages from scratch instead of fixing Stage-1 (ablation) | SCAN Stage-2 128-256 · Cityscapes Photo->Labels | pixel accuracy: 0.72 -> 0.36 | Hurt |
| fix Stage-1 weights rather than fine-tuning it (ablation) | SCAN Stage-2 128-256 · Cityscapes Photo->Labels | pixel accuracy: 0.62 -> 0.72 | Improved |
| use learned per-pixel weight fusion instead of uniform weight (ablation) | SCAN 128-256 · Cityscapes Labels->Photo | pixel accuracy: 0.59 -> 0.64 | Improved |
| use learned per-pixel weight fusion instead of residual fusion (ablation) | SCAN 128-256 · Cityscapes Photo->Labels | pixel accuracy: 0.68 -> 0.72 | Improved |
| add adaptive fusion block (ablation) | SCAN Stage-2 128-256 · Cityscapes Labels->Photo (ablation) | pixel accuracy: 0.613 -> 0.637 | Improved |
| add skip connection from the input (ablation) | SCAN Stage-2 128-256 · Cityscapes Labels->Photo (ablation) | pixel accuracy: 0.593 -> 0.637 | Improved |
| add both skip connection and fusion block (ablation) | SCAN Stage-2 128-256 · Cityscapes Labels->Photo (ablation) | pixel accuracy: 0.513 -> 0.637 | Improved |
| replace CycleGAN with SCAN on Aerial->Map | CycleGAN · Aerial->Map | PSNR: 21.59 -> 25.15 | Improved |
Closest papers by meaning
- Unpaired Image-To-Image Translation Using Cycle-Consistent Adversarial Networks (ICCV 2017 · similarity 0.91)
- Unsupervised Image-to-Image Translation Networks (NeurIPS 2017 · similarity 0.91)
- On Translation and Reconstruction Guarantees of the Cycle-Consistent Generative Adversarial Networks (NeurIPS 2022 · similarity 0.91)
- TraVeLGAN: Image-To-Image Translation by Transformation Vector Learning (CVPR 2019 · similarity 0.90)
- Few-Shot Unsupervised Image-to-Image Translation (ICCV 2019 · similarity 0.90)
- Adversarial Self-Defense for Cycle-Consistent GANs (NeurIPS 2019 · similarity 0.90)
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.
