Research map / Image generation / GANs
LARGE: Latent-Based Regression Through GAN Semantics
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 |
|---|---|---|---|
| Use GAN latent semantic distances as the regression feature instead of generating labeled training images (ablation) | LARGE head-pose estimation · CelebA-HQ test set | MAE: 3.23 ± 3.67 -> 2.97 ± 2.76 | Helped |
| Use e4e inversion rather than pSp-based inversion (ablation) | LARGE · CelebA-HQ pose and age estimation | MAE | Helped |
| Use semantic-preserving e4e inversion rather than reconstruction-focused pSp inversion (ablation) | LARGE · CelebA-HQ pose and age estimation | MAE | Helped |
| Replace the unsupervised per-layer weighted distance with Euclidean distance (ablation) | LARGE W+ to W boundary mapping · Pose and age regression | MAE | Hurt |
| Replace the unsupervised per-layer weighted distance with a single layer distance (ablation) | LARGE W+ to W boundary mapping · Pose and age regression | MAE | Hurt |
| Use all 18 per-layer distances as regression features instead of the weighted distance (ablation) | LARGE W+ to W boundary mapping · Pose and age regression | MAE | Mixed |
| Use polynomial regression of degrees 2, 3, or 5 instead of linear calibration (ablation) | LARGE · CelebA-HQ pose and CACD age estimation | MAE | Hurt |
| Change linear-regression regularization among none, L1, L2, and ElasticNet (ablation) | LARGE calibration · Few-shot regression | MAE | No clear effect |
| Use GAN-latent sorting scores instead of CLIP embedding distances | LARGE text-guided face sorting · Hair color, makeup, expression, and hair length | preferred responses: 73.55% vs 25.81%; 70.53% vs 18.25%; 53.45% vs 43.27%; 84.73% vs 2.18% | Helped |
| Use GAN-latent sorting scores instead of SSV pose predictions | LARGE cat sorting · Cat yaw and pitch | preferred responses: 66.67% vs 28.99%; 82.50% vs 7.50% | Helped |
| Use StyleGAN latent distances rather than alternative feature-space SVM distances | LARGE · CelebA-HQ pose and age regression | MAE | Helped |
| Use a text-derived StyleCLIP semantic boundary for age regression | LARGE · CACD age estimation | MAE (years): 7.46 vs 7.59 | Helped |
Closest papers by meaning
- Using latent space regression to analyze and leverage compositionality in GANs (ICLR 2021 · similarity 0.93)
- Closed-Form Factorization of Latent Semantics in GANs (CVPR 2021 · similarity 0.91)
- FEditNet: Few-Shot Editing of Latent Semantics in GAN Spaces (AAAI 2023 · similarity 0.91)
- Unsupervised Discovery of Interpretable Directions in the GAN Latent Space (ICML 2020 · similarity 0.90)
- E$^2$GAN: Efficient Training of Efficient GANs for Image-to-Image Translation (ICML 2024 · similarity 0.89)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · 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.