Research map / Image generation / GANs
Distilling Portable Generative Adversarial Networks for Image Translation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 the full generator-and-discriminator distillation method instead of a half-channel student trained from scratch | pix2pix U-net student · Cityscapes | Per-pixel accuracy: 51.62 -> 52.22 | Helped |
| Use the full method instead of vanilla pixel-loss distillation with half channels | pix2pix U-net student · Cityscapes | Per-pixel accuracy: 50.42 -> 52.22 | Helped |
| Use vanilla pixel-loss distillation instead of training the half-channel student from scratch | pix2pix U-net student · Cityscapes | Per-pixel accuracy: 51.62 -> 50.42 | Hurt |
| Add perceptual loss to the baseline student (ablation) | pix2pix U-net student · Cityscapes | IOU: 7.93 -> 8.01 | Helped |
| Add L1 loss to perceptual loss (ablation) | pix2pix U-net student · Cityscapes | Per-pixel accuracy: 51.22 -> 51.82 | Helped |
| Use the combined L1 and perceptual losses instead of the baseline student (ablation) | pix2pix U-net student · Cityscapes | IOU: 7.93 -> 8.06 | Helped |
| Add triplet loss to the teacher-generated samples loss (ablation) | pix2pix U-net student discriminator · Cityscapes | Per-pixel accuracy: 51.66 -> 52.05 | Helped |
| Use teacher-generated samples loss with triplet loss instead of the baseline student (ablation) | pix2pix U-net student discriminator · Cityscapes | IOU: 7.93 -> 8.08 | Helped |
| Combine generator and discriminator distillation losses (ablation) | pix2pix U-net student · Cityscapes | Per-pixel accuracy: 51.82 -> 52.22 | Helped |
| Reduce generator channels to half or one-quarter and apply the proposed distillation | pix2pix U-net student · Cityscapes | Per-pixel accuracy: 52.22; 51.57 | Mixed |
| Apply vanilla pixel matching to a portable CycleGAN student instead of the proposed method | CycleGAN student · Horse→zebra | image quality | Hurt |
| Distill the student generator and discriminator from the teacher GAN | CycleGAN student · Summer→winter | image quality | Helped |
Closest papers by meaning
- Adversarial Learning of Portable Student Networks (AAAI 2018 · similarity 0.91)
- Wavelet Knowledge Distillation: Towards Efficient Image-to-Image Translation (CVPR 2022 · similarity 0.91)
- Teachers Do More Than Teach: Compressing Image-to-Image Models (CVPR 2021 · similarity 0.90)
- Distilling Representations from GAN Generator via Squeeze and Span (NeurIPS 2022 · similarity 0.90)
- Compressing Image-to-Image Translation GANs Using Local Density Structures on Their Learned Manifold (AAAI 2024 · similarity 0.88)
- E$^2$GAN: Efficient Training of Efficient GANs for Image-to-Image Translation (ICML 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.