Research map / Image generation / GANs
Semantically Decomposing the Latent Spaces of Generative Adversarial Networks
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 |
|---|---|---|---|
| use the pairwise identity/observation latent decomposition | GANs on MS-Celeb-1M · MS-Celeb-1M test pairs | ID-Div: .497 -> .588 | Helped |
| use the pairwise identity/observation latent decomposition | GANs on MS-Celeb-1M · MS-Celeb-1M test pairs | FaceNet accuracy: .851 -> .857 | Mixed |
| use k=4 images per identity instead of pairwise training (ablation) | SD-DCGAN · MS-Celeb-1M test pairs | FaceNet accuracy: .749 -> .776 | Helped |
| use k=4 images per identity instead of pairwise training (ablation) | SD-DCGAN · MS-Celeb-1M human judgments | accuracy: .723 -> .688 | Hurt |
| set identity latent dimension to 25 instead of 50 (ablation) | SD-DCGAN · MS-Celeb-1M test pairs | FaceNet accuracy: .749 -> .764 | Helped |
| set identity latent dimension to 75 instead of 50 (ablation) | SD-DCGAN · MS-Celeb-1M test pairs | FaceNet accuracy: .749 -> .743 | Hurt |
| replace the Siamese discriminator with a stacked-channel discriminator (ablation) | SD-DCGAN · MS-Celeb-1M test pairs | FaceNet accuracy: .749 -> .757 | No clear effect |
| use BEGAN rather than DCGAN as the GAN architecture (ablation) | SD-GAN · MS-Celeb-1M test pairs | FaceNet accuracy: .749 -> .857 | Helped |
| use BEGAN rather than DCGAN as the GAN architecture (ablation) | SD-GAN · MS-Celeb-1M test pairs | ID-Div: .521 -> .588 | Helped |
| use a one-hot identity embedding rather than random identity codes | AC-DCGAN · MS-Celeb-1M test pairs | ID-Div: .588 -> .497 | Hurt |
| train the modified DR-GAN to disentangle identity using identity classification (ablation) | DR-GAN · MS-Celeb-1M | observation diversity | Hurt |
| use k=4 in SD-BEGAN (ablation) | SD-BEGAN · MS-Celeb-1M | mode collapse | Hurt |
Closest papers by meaning
- Interpreting the Latent Space of GANs for Semantic Face Editing (CVPR 2020 · similarity 0.88)
- DeLiGAN : Generative Adversarial Networks for Diverse and Limited Data (CVPR 2017 · similarity 0.87)
- Adversarial Latent Autoencoders (CVPR 2020 · similarity 0.87)
- Ensembling With Deep Generative Views (CVPR 2021 · similarity 0.87)
- CapHuman: Capture Your Moments in Parallel Universes (CVPR 2024 · similarity 0.86)
- Generative Unlearning for Any Identity (CVPR 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.