Research map / Image generation / GANs
Sequential Training of GANs Against GAN-Classifiers Reveals Correlated "Knowledge Gaps" Present Among Independently Trained GAN Instances
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 |
|---|---|---|---|
| Train SG2 to fool one ResNet-50 classifier | StyleGAN2 on FFHQ · FFHQ, unseen ResNet-50 classifiers | accuracy on generated images: 0.993±0.002 -> 0.05±0.02 | Helped |
| Train SG2 to fool an Inception-v3 classifier | StyleGAN2 on FFHQ · FFHQ, unseen Inception-v3 classifiers | accuracy on generated images: 0.16±0.26 | Helped |
| Train SG2 to fool one MobileNetV2 instance | StyleGAN2 on FFHQ · FFHQ, unseen MobileNetV2 classifiers | accuracy on generated images: 0.41±0.38 | Mixed |
| Train SG2 to fool multiple MobileNetV2 instances instead of one (ablation) | StyleGAN2 on FFHQ · FFHQ, unseen MobileNetV2 classifiers | accuracy on generated images: 0.98 -> 0.04 | Helped |
| Add the classifier-augmented fool-all loss across sequential rounds | StyleGAN2 on FFHQ · FFHQ, FID across iterations 0–4 | FID: 36.98, 36.62, 36.67, 36.39, 36.83 | No clear effect |
| Train later-iteration generators to fool classifiers from preceding iterations | StyleGAN2 on FFHQ · FFHQ, classifiers tested on later-iteration generators | accuracy: 0.993±0.002 -> 0.028±0.004 | Hurt |
| Use the memoryless loss, fooling only the immediately preceding classifier (ablation) | StyleGAN2 on FFHQ · FFHQ, classifiers from two iterations earlier | detectability | Mixed |
| Train a DCGAN to fool a classifier | DCGAN on MNIST · MNIST, held-out classifiers and image quality | fooling ability | Mixed |
| Use truncated samples from one SG2 generator to train the classifier (ablation) | StyleGAN2 on FFHQ · FFHQ, truncated samples from an independent generator | classification accuracy | Helped |
| Apply truncation with ψ=0.5 (ablation) | StyleGAN2 on FFHQ · FFHQ | FID: ~37 -> ~81 | Hurt |
| Change the GAN-classifier architecture (ablation) | StyleGAN2 on FFHQ · FFHQ, cross-architecture classifier generalization | accuracy on generated images: 0.05±0.02; 0.74±0.12; 0.5±0.39 | Mixed |
| Finetune an iteration 0 GAN against a pre-trained detector instead of training from scratch (ablation) | StyleGAN2 on FFHQ · FFHQ, next-iteration artifact behavior | detector behavior | No clear effect |
Closest papers by meaning
- Improved Techniques for Training GANs (NeurIPS 2016 · similarity 0.91)
- Training Generative Adversarial Networks with Limited Data (NeurIPS 2020 · similarity 0.90)
- Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions (CVPR 2024 · similarity 0.90)
- Projected GANs Converge Faster (NeurIPS 2021 · similarity 0.90)
- The GAN is dead; long live the GAN! A Modern GAN Baseline (NeurIPS 2024 · similarity 0.89)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.89)
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.