Research map / Image generation / GANs
Image Manipulation with Perceptual Discriminators
Diana Sungatullina, Egor Zakharov, Dmitry Ulyanov, Victor Lempitsky
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 perceptual VGG features to the discriminator (VGG-GAN vs DCGAN) | DCGAN · CelebA C2ST smile | C2ST cross-entropy x10^-2: 0.6 -> 8.6 | Improved |
| add perceptual VGG features to the discriminator (VGG-GAN vs DCGAN) | DCGAN · CelebA C2ST hair color | C2ST cross-entropy x10^-2: 0.06 -> 0.96 | Improved |
| add perceptual VGG features to the discriminator (VGG-GAN vs DCGAN) | DCGAN · CelebA classification loss gender | log-loss: 1.5 -> 0.1 | Improved |
| add perceptual VGG features to the discriminator (VGG-GAN vs DCGAN) | DCGAN · CelebA classification loss hair color | log-loss: 2.33 -> 1.3 | Improved |
| use VGG* (leaky ReLU, average pooling) instead of VGG as feature network (ablation) | VGG-GAN · CelebA C2ST hair color | C2ST cross-entropy x10^-2: 0.96 -> 1.29 | Improved |
| use VGG* (leaky ReLU, average pooling) instead of VGG as feature network (ablation) | VGG-GAN · CelebA C2ST smile | C2ST cross-entropy x10^-2: 8.6 -> 5.2 | Hurt |
| use VGG* instead of VGG as feature network (ablation) | VGG-GAN · CelebA qualitative | image quality | No clear effect |
| pretrained weights instead of random weights in the perceptual feature network (ablation) | MS-CycleGAN on CelebA-HQ 256 · CelebA-HQ male-female | qualitative realism | Improved |
| perceptual discriminator (VGG*-MS-CycleGAN) vs CycleGAN | CycleGAN · apple to orange | qualitative realism | Improved |
| perceptual discriminator vs CycleGAN | CycleGAN · Monet-photo style transfer | qualitative realism | No clear effect |
| perceptual discriminator vs additive perceptual loss | standard discriminator · aligned image translation | image quality | No clear effect |
| naive perceptual discriminator | unconditional generation · unconditional image generation | training stability | Hurt |
Closest papers by meaning
- Reusing Discriminators for Encoding: Towards Unsupervised Image-to-Image Translation (CVPR 2020 · similarity 0.88)
- Unpaired Image-To-Image Translation Using Cycle-Consistent Adversarial Networks (ICCV 2017 · similarity 0.88)
- Discriminative Region Proposal Adversarial Networks for High-Quality Image-to-Image Translation (ECCV 2018 · similarity 0.88)
- Discriminative Region Proposal Adversarial Network for High-Quality Image-to-Image Translation (IJCV 2019 · similarity 0.87)
- Lightweight Generative Adversarial Networks for Text-Guided Image Manipulation (NeurIPS 2020 · similarity 0.87)
- Few-Shot Unsupervised Image-to-Image Translation (ICCV 2019 · similarity 0.87)
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.
