Research map / Label-efficient and robust learning / Domain adaptation and generalization
Augmented CycleGAN: Learning Many-to-Many Mappings from Unpaired Data
Amjad Almahairi, Sai Rajeshwar, Alessandro Sordoni, Philip Bachman, Aaron Courville
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 |
|---|---|---|---|
| Use AugCGAN instead of CycleGAN | Edges-to-photos shoe translation · Shoes test set | Avg. L1 reconstruction error: 0.1837 -> 0.0698 | Improved |
| Use AugCGAN instead of StochCGAN | Edges-to-photos shoe translation · Shoes test set | Avg. L1 reconstruction error: 0.0794 -> 0.0698 | Improved |
| Add 10% paired data | AugCGAN · Shoes test set, edges-to-photos | Avg. L1 reconstruction error: 0.0698 -> 0.0562 | Improved |
| Use AugCGAN instead of Δ-GAN† | Shoes translation, photos-to-edges · Shoes test set | MSE: 0.0096 -> 0.0079 | Improved |
| Add 10% paired data | AugCGAN · Shoes test set, photos-to-edges | MSE: 0.0079 -> 0.0052 | Improved |
| Inject latent noise only in the last 3 layers rather than throughout the generator (ablation) | AugCGAN · CelebA male-to-female generation | LPIPS diversity: 0.108 +/- 0.003 -> 0.059 +/- 0.001 | Hurt |
| Replace AugCGAN with StochCGAN | CelebA male-to-female generation · CelebA | LPIPS diversity: 0.108 +/- 0.003 -> 0.008 +/- 0.000 | Hurt |
| Use AugCGAN instead of the baseline classifier | Semi-supervised attribute prediction · CelebA, s = 1% | P@10 / NDCG@10: 63.36 / 79.25 -> 64.38 / 80.59 | Improved |
| Use AugCGAN instead of the baseline classifier | Semi-supervised attribute prediction · CelebA, s = 10% | P@10 / NDCG@10: 67.34 / 84.21 -> 68.83 / 85.51 | Improved |
| Use augmented latent spaces and encoders rather than stochastic CycleGAN's direct stochastic cycle | Stochastic cross-domain mapping · Generated edges used as input for shoe generation | Conditional output diversity | Improved |
Closest papers by meaning
- Guiding the One-to-One Mapping in CycleGAN via Optimal Transport (AAAI 2019 · similarity 0.87)
- Variational Inference for Cyclic Learning (ICLR 2026 · similarity 0.85)
- Kernel of CycleGAN as a principal homogeneous space (ICLR 2020 · similarity 0.85)
- Increasing the Robustness of Semantic Segmentation Models with Painting-by-Numbers (ECCV 2020 · similarity 0.85)
- Training-Free Dataset Pruning for Instance Segmentation (ICLR 2025 · similarity 0.84)
- A Generative Approach at the Instance-Level for Image Segmentation Under Limited Training Data Conditions (Student Abstract) (AAAI 2025 · similarity 0.83)
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.
