Research map / Image generation / GANs
Disentangled Inference for GANs With Latently Invertible Autoencoder
Jiapeng Zhu, Deli Zhao, Bo Zhang, Bolei Zhou
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 y-space (after mapping network) instead of z-space for interpolation | StyleGAN · FFHQ | perceptual path length: 40.25 -> 23.26 | Improved |
| add invertible network in place of the MLP mapping (LIA/GAN vs VAE/GAN) (ablation) | StyleGAN encoder · FFHQ reconstruction | FID: 22.26 -> 19.26 | Improved |
| add invertible network in place of the MLP mapping (LIA/GAN vs VAE/GAN) (ablation) | StyleGAN encoder · FFHQ reconstruction | MSE: 23.18 -> 11.79 | Improved |
| add invertible network in place of the MLP mapping (LIA/GAN vs VAE/GAN) (ablation) | StyleGAN encoder · FFHQ reconstruction | SWD: 15.82 -> 12.16 | Improved |
| use LIA encoder instead of MSE-based GAN inversion | StyleGAN · FFHQ reconstruction | FID: 44.79 -> 19.26 | Improved |
| use LIA encoder instead of ALI adversarial inference | StyleGAN · FFHQ reconstruction | MSE: 32.61 -> 11.79 | Improved |
| initialize finetuning from LIA encoder output instead of average y (ablation) | latent finetuning · FFHQ last 1,000 images | FID: 57.09 -> 39.52 | Improved |
| initialize finetuning from LIA encoder output instead of random sampling (ablation) | latent finetuning · FFHQ last 1,000 images | FID: 105.52 -> 39.52 | Improved |
| initialize finetuning from LIA encoder output instead of average y (ablation) | latent finetuning · FFHQ last 1,000 images | PSNR: 19.33 -> 23.51 | Improved |
| optimize in disentangled y-space instead of z-space (encoder init) (ablation) | latent finetuning · Web images (100) | MSE: 0.0319 -> 0.0055 | Improved |
| optimize in y-space instead of z-space (random init) (ablation) | latent finetuning · Web images (100) | FID: 216.8 -> 75.34 | Improved |
| optimize in y-space instead of z-space (average init) (ablation) | latent finetuning · Web images (100) | MSE: 0.058 -> 0.0075 | Improved |
Closest papers by meaning
- Interpreting the Latent Space of GANs for Semantic Face Editing (CVPR 2020 · similarity 0.91)
- In-Domain GAN Inversion for Faithful Reconstruction and Editability (TPAMI 2024 · similarity 0.91)
- In-Domain GAN Inversion for Real Image Editing (ECCV 2020 · similarity 0.91)
- InfoGAN-CR and ModelCentrality: Self-supervised Model Training and Selection for Disentangling GANs (ICML 2020 · similarity 0.90)
- Adversarial Latent Autoencoders (CVPR 2020 · similarity 0.89)
- Complexity Matters: Rethinking the Latent Space for Generative Modeling (NeurIPS 2023 · 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.
