Research map / Image generation / GANs
Dual Projection Generative Adversarial Networks for Conditional Image Generation
Ligong Han, Martin Renqiang Min, Anastasis Stathopoulos, Yu Tian, Ruijiang Gao, Asim Kadav, Dimitris N. Metaxas
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 dual projection (untied class embeddings with classification losses) to Proj-GAN (ablation) | Proj-GAN · VGGFace200 | FID: 61.27 -> 23.15 | Improved |
| add dual projection to Proj-GAN (ablation) | Proj-GAN · ImageNet 64x64 (60k iters) | FID: 24.57 -> 21.46 | Improved |
| add dual projection to Proj-GAN (ablation) | Proj-GAN · CIFAR100 | FID: 10.02 -> 10.55 | Hurt |
| add the two cross-entropy label-matching losses (lambda > 0 instead of lambda = 0) (ablation) | P2GAN · VGGFace200 | FID: 62.46 -> 23.15 | Improved |
| keep the psi residual term (do not set psi to zero) (ablation) | P2GAN · VGGFace200 | FID: 87.01 -> 23.15 | Improved |
| use dual projection instead of naive hybrid of Proj-GAN and f-cGAN losses (ablation) | P2GAN · VGGFace200 | FID: 32.31 -> 23.15 | Improved |
| add amortized per-sample weights with penalty (P2GAN-ap) to P2GAN | P2GAN · VGGFace200 | FID: 23.15 -> 18.99 | Improved |
| add amortized per-sample weights with penalty (P2GAN-ap) to P2GAN | P2GAN · ImageNet 64x64 (80k iters) | FID: 19.22 -> 18.31 | Improved |
| add amortized per-sample weights with penalty (P2GAN-ap) to P2GAN | P2GAN · CIFAR100 (62k iters) | FID: 10.55 -> 9.82 | Improved |
| add a penalty term on lambda to scalar-weighted P2GAN (P2GAN-s to P2GAN-sp) (ablation) | P2GAN-s · ImageNet 64x64 (80k iters) | FID: 18.62 -> 20.68 (ImageNet), 9.03 -> 9.51 (CIFAR100), 20.59 -> 20.18 (VGGFace200) | Mixed |
| use the alternative weighting (fix weight of L_D to 1) for scalar weighting (ablation) | P2GAN-s · CIFAR100 and VGGFace200 | FID: 9.03 -> 9.49 (CIFAR100), 20.59 -> 20.21 (VGGFace200) | No clear effect |
| add adaptive weights (P2GAN-w) to P2GAN | P2GAN · VGGFace200 | max-FID: 209.86 -> 127.43 | Improved |
Closest papers by meaning
- CcGAN: Continuous Conditional Generative Adversarial Networks for Image Generation (ICLR 2021 · similarity 0.93)
- Continuous Conditional Generative Adversarial Networks: Novel Empirical Losses and Label Input Mechanisms (TPAMI 2022 · similarity 0.93)
- ContraGAN: Contrastive Learning for Conditional Image Generation (NeurIPS 2020 · similarity 0.92)
- A Unified View of cGANs with and without Classifiers (NeurIPS 2021 · similarity 0.92)
- Turning Waste into Wealth: Leveraging Low-Quality Samples for Enhancing Continuous Conditional Generative Adversarial Networks (AAAI 2024 · similarity 0.92)
- Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions (CVPR 2024 · 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.
