Research map / Image generation / GANs
Discriminator Feature-Based Inference by Recycling the Discriminator of GANs
Duhyeon Bang, Seoungyoon Kang, Hyunjung Shim
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 DFI (discriminator features + latent reconstruction) instead of VAE | inference mapping · CIFAR10 | FID: 234.03 -> 31.54 | Improved |
| use DFI (WGAN-GP) instead of ALICE | inference mapping · CelebA | FID: 16.36 -> 8.73 | Improved |
| use DFI (WGAN-GP) instead of ALI/BiGAN | inference mapping · Fashion MNIST | FID: 69.99 -> 11.85 | Improved |
| use DFI with SNGAN base instead of other GANs | DFI · CIFAR10 | FID: 31.54 (WGAN-GP) vs 61.88 (SNGAN) | Hurt |
| use LSGAN base instead of DCGAN for DFI | DFI · CelebA | FID: 12.86 -> 24.99 | Hurt |
| add latent reconstruction loss instead of image reconstruction loss (ablation) | DFI · CelebA | FID / LPIPS | Mixed |
| add latent optimization (50 iterations) after inference mapping (ablation) | DFI · CelebA | user study / color fidelity | Improved |
| use discriminator features instead of pre-trained VGG16 Pool5 features (ablation) | DFI · CelebA | LPIPS, FID, user study | Improved |
| add GAP of intermediate layer Actv64-1 to discriminator features (ablation) | DFI (SNGAN) · CelebA | FID: 8.91 -> 8.54 | Improved |
| add GAP of Actv128-0 to the base discriminator feature (ablation) | DFI (SNGAN) · CelebA | FID: 8.91 -> 10.34 | Hurt |
| combine GAP features from all three intermediate layers (ablation) | DFI (SNGAN) · CelebA | LPIPS / FID: LPIPS 0.1931 -> 0.1887; FID 8.91 -> 10.32 | Mixed |
| use 4x4 average pooling instead of GAP (1x1) on Actv64-1 (ablation) | DFI (SNGAN) · CelebA | LPIPS / FID: LPIPS 0.1903 -> 0.1835; FID 8.54 -> 8.26 | Improved |
Closest papers by meaning
- Generalized Adversarially Learned Inference (AAAI 2021 · similarity 0.87)
- Adversarial Feedback Loop (ICCV 2019 · similarity 0.87)
- cGANs with Projection Discriminator (ICLR 2018 · similarity 0.87)
- Direct Discriminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminator (ICML 2025 · similarity 0.86)
- Your GAN is Secretly an Energy-based Model and You Should Use Discriminator Driven Latent Sampling (NeurIPS 2020 · similarity 0.86)
- Discriminative Forests Improve Generative Diversity for Generative Adversarial Networks (AAAI 2024 · similarity 0.86)
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.
