Research map / Image generation / GANs
Using latent space regression to analyze and leverage compositionality in GANs
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 |
|---|---|---|---|
| provide a binary known/missing-pixel mask to the encoder (ablation) | latent regression with a pretrained GAN · ProGAN Living Room collage | FID: 19.50 -> 14.90 | Helped |
| provide a binary known/missing-pixel mask to the encoder (ablation) | latent regression with a pretrained GAN · ProGAN CelebA-HQ collage | FID: 16.24 -> 17.41 | Hurt |
| provide a binary known/missing-pixel mask to the encoder (ablation) | latent regression with a pretrained GAN · StyleGAN Church, Car, and FFHQ collages | FID: 6.13 -> 6.09; 6.23 -> 5.38; 26.25 -> 24.09 | Helped |
| provide a binary known/missing-pixel mask to the encoder (ablation) | latent regression with a pretrained GAN · ProGAN Church collage | Masked L1: 0.26 -> 0.26 | No clear effect |
| remove latent recovery loss (ablation) | ProGAN Church encoder training · masked reconstruction | LPIPS(alex): 0.291 -> 0.305 | Hurt |
| remove latent recovery loss (ablation) | ProGAN Church encoder training · masked reconstruction | L1: 0.194 -> 0.194 | No clear effect |
| remove perceptual loss (ablation) | ProGAN Church encoder training · masked reconstruction | LPIPS(alex): 0.291 -> 0.318 | Hurt |
| remove perceptual loss (ablation) | ProGAN Church encoder training · masked reconstruction | L1: 0.194 -> 0.197 | Hurt |
| use StyleGAN rather than ProGAN | GAN generator · face part independence | Average variation outside replaced part: 0.093 -> 0.105 | Mixed |
| compose by encoding a masked pixel collage rather than interpolating latent or pixel codes | latent regression image editing · living room and church edits | masked L1 distance to context, target, and collage | Helped |
| train the encoder with masked inputs and a mask channel (ablation) | latent regression with a pretrained GAN · completion of masked real images | visual reconstruction quality | Helped |
Closest papers by meaning
- LARGE: Latent-Based Regression Through GAN Semantics (CVPR 2022 · similarity 0.93)
- Low-Rank Subspaces in GANs (NeurIPS 2021 · similarity 0.91)
- Complexity Matters: Rethinking the Latent Space for Generative Modeling (NeurIPS 2023 · similarity 0.91)
- Enjoy Your Editing: Controllable GANs for Image Editing via Latent Space Navigation (ICLR 2021 · similarity 0.90)
- Discovering Interpretable Latent Space Directions of GANs Beyond Binary Attributes (CVPR 2021 · similarity 0.90)
- WeditGAN: Few-Shot Image Generation via Latent Space Relocation (AAAI 2024 · 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.