Research map / Image generation / GANs
Learning Energy-Based Models with Adversarial Training
Xuwang Yin, Shiying Li, Gustavo K. Rohde
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 a more diverse p0 dataset (ablation) | binary AT generative model · CIFAR-10 | FID | Improved |
| increase the number of PGD steps K during training (ablation) | binary AT generative model · CIFAR-10 | FID | Improved |
| progressively increase K from 0 to 25 instead of fixed K=25 (ablation) | binary AT generative model · CIFAR-10 | training convergence speed | Improved |
| pretrain the discriminator on ImageNet classification (ablation) | binary AT generative model · CelebA-HQ 256 | FID | Improved |
| add R1 regularization (30) when using a pretrained model (ablation) | binary AT generative model · CelebA-HQ 256 | FID | Improved |
| add R1 regularization when training from scratch (ablation) | binary AT generative model · CelebA-HQ 256 | FID | No clear effect |
| vary number of sampling steps with fixed total perturbation (64x0.1, 32x0.2, 16x0.4) (ablation) | binary AT generative model · CIFAR-10 | FID: 13.07, 13.21, 13.49 | No clear effect |
| vary number of sampling steps with fixed total perturbation (40x4.0, 20x8.0, 10x16.0) (ablation) | binary AT generative model · CelebA-HQ 256 | FID: 19.19, 18.97, 19.19 | No clear effect |
| apply Gaussian smoothing (sigma=10) to source p0 images (ablation) | binary AT generative model · CelebA-HQ 256 | visual artifacts and FID | Mixed |
| use the EBM likelihood gradient with the PGD sampler instead of the binary AT objective (ablation) | binary AT generative model · training stability | training stability | Hurt |
| binary AT versus explicit EBMs and AT-based models | binary AT generative model · CIFAR-10 | FID: 38.2 -> 13.21 | Improved |
| binary AT versus VAEBM on CelebA-HQ 256 | binary AT generative model · CelebA-HQ 256 | FID: 20.38 -> 17.31 | Improved |
Closest papers by meaning
- Improving Adversarial Energy-Based Model via Diffusion Process (ICML 2024 · similarity 0.92)
- A Unified Contrastive Energy-based Model for Understanding the Generative Ability of Adversarial Training (ICLR 2022 · similarity 0.91)
- Scalable Energy-Based Models via Adversarial Training: Unifying Discrimination and Generation (ICLR 2026 · similarity 0.91)
- Calibrating Energy-based Generative Adversarial Networks (ICLR 2017 · similarity 0.91)
- Energy-based Generative Adversarial Networks (ICLR 2017 · similarity 0.90)
- Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood (ICLR 2024 · 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.
