Research map / Image generation / GANs
Ensembling With Deep Generative Views
Check 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 |
|---|---|---|---|
| Add fine-layer style-mixing GAN views at test time | Car classifier trained on real images · Car classification | accuracy: 96.6% -> 96.9% | Helped |
| Finetune the classifier on GAN reconstructions or perturbed reconstructions, then ensemble GAN views | Car classifier · Car classification | accuracy: 96.9% -> 97.9% | Helped |
| Add GAN views at test time without GAN-domain classifier training | Cat classifier trained on real images · Cat classification | accuracy | No clear effect |
| Train with perturbed GAN reconstructions | Cat classifier · Cat classification | accuracy: 92.3% -> 93.2% | Helped |
| Add GAN-view ensembling after training with GAN perturbations | Cat classifier · Cat classification | accuracy: 93.2% -> 93.3% | Helped |
| Combine fine-layer style-mixing with color and spatial jitter | CelebA-HQ attribute classifiers trained on images · CelebA-HQ, 40 attributes | accuracy: 93.59 -> 93.89 (Smiling); -0.52 (Big Lips) | Mixed |
| Use fine-layer style-mixing instead of other latent perturbations (ablation) | CelebA-HQ attribute classifiers · CelebA-HQ, 12-attribute subset | accuracy | Helped |
| Combine GAN style-mixing views with conventional color and spatial views (ablation) | CelebA-HQ attribute classifiers · CelebA-HQ, 40 attributes | accuracy | Helped |
| Increase the number of GAN views in the ensemble (ablation) | CelebA-HQ Smiling classifier · CelebA-HQ Smiling | accuracy | Helped |
| Use GAN reconstructions to train the classifier (ablation) | CelebA-HQ attribute classifiers · CelebA-HQ, 40 attributes | accuracy | Helped |
| Train the classifier using fine-layer style-mixing samples (ablation) | CelebA-HQ attribute classifiers · CelebA-HQ, 40 attributes | accuracy | Mixed |
| Increase latent optimization from 0 to 250 steps (ablation) | CelebA-HQ Smiling GAN reconstruction · CelebA-HQ Smiling | accuracy: 92.2 -> 93.6 | Helped |
Closest papers by meaning
- Interpreting the Latent Space of GANs for Semantic Face Editing (CVPR 2020 · similarity 0.89)
- Adversarial Latent Autoencoders (CVPR 2020 · similarity 0.89)
- Generative Hierarchical Features From Synthesizing Images (CVPR 2021 · similarity 0.88)
- Navigating the GAN Parameter Space for Semantic Image Editing (CVPR 2021 · similarity 0.88)
- LARGE: Latent-Based Regression Through GAN Semantics (CVPR 2022 · similarity 0.88)
- Editing in Style: Uncovering the Local Semantics of GANs (CVPR 2020 · similarity 0.88)
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.