Research map / Label-efficient and robust learning / Self-supervised and contrastive
Unsupervised Discovery of Interpretable Directions in the GAN Latent Space
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 |
|---|---|---|---|
| Learn directions jointly to make transformations distinguishable | Unsupervised GAN latent direction discovery · MNIST, Anime, CelebA, BigGAN | RCA: MNIST: 0.48 -> 0.88; Anime: 0.89 -> 0.99; CelebA: 0.82 -> 0.9; BigGAN: 0.66 -> 0.85 | Helped |
| Learn directions jointly to make transformations distinguishable | Unsupervised GAN latent direction discovery · MNIST, Anime, BigGAN | MOS: MNIST: 0.06 -> 0.47; Anime: 0.00 -> 0.26; BigGAN: 0.66 -> 0.69 | Helped |
| Learn directions jointly to make transformations distinguishable | Unsupervised GAN latent direction discovery · CelebA | MOS: 0.29 -> 0.30 | No clear effect |
| Use the BigGAN background-removal direction to create synthetic segmentation masks | Weakly supervised saliency detection · ECSSD test | MAE: 0.099 | Helped |
| Set K to 32 rather than K = 16 (ablation) | Spectral Norm GAN direction discovery · MNIST | MOS: 0.5 -> 0.58 | Helped |
| Increase K from 32 to 64 (ablation) | Spectral Norm GAN direction discovery · MNIST | MOS: 0.58 -> 0.47 | Hurt |
| Increase K from 64 to 128 (ablation) | Spectral Norm GAN direction discovery · MNIST | MOS: 0.47 -> 0.46 | Hurt |
| Increase K from 15 to 60 (ablation) | BigGAN direction discovery · BigGAN | MOS: 0.3 -> 0.38 | Helped |
| Increase K from 60 to 90 (ablation) | BigGAN direction discovery · BigGAN | MOS: 0.38 -> 0.75 | Helped |
| Increase K from 90 to 120 (ablation) | BigGAN direction discovery · BigGAN | MOS: 0.75 -> 0.69 | Hurt |
| Use a small direction count (ablation) | GAN direction discovery · MNIST and BigGAN | MOS: MNIST: 0.58 at K = 32; BigGAN: 0.3 at K = 15 and K = 30 | Mixed |
| Set shift-loss multiplier lambda to 0 instead of 0.25 (ablation) | Spectral Norm GAN direction discovery · MNIST | MOS: 0.47 -> 0.27 | Hurt |
Closest papers by meaning
- LARGE: Latent-Based Regression Through GAN Semantics (CVPR 2022 · similarity 0.90)
- Discovering Interpretable Latent Space Directions of GANs Beyond Binary Attributes (CVPR 2021 · similarity 0.89)
- Do Not Escape From the Manifold: Discovering the Local Coordinates on the Latent Space of GANs (ICLR 2022 · similarity 0.89)
- PandA: Unsupervised Learning of Parts and Appearances in the Feature Maps of GANs (ICLR 2023 · similarity 0.89)
- Self-Supervised Enhancement of Latent Discovery in GANs (AAAI 2022 · similarity 0.89)
- NoiseCLR: A Contrastive Learning Approach for Unsupervised Discovery of Interpretable Directions in Diffusion Models (CVPR 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.