Research map / Image generation / GANs
Linearly Controllable GAN: Unsupervised Feature Categorization and Decomposition for Image Generation and Manipulation
sehyung lee, Mijung Kim, Yeongnam Chae, Bjorn Stenger
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 |
|---|---|---|---|
| add contrastive loss and spectral regularization to generator training (ablation) | LC-GAN · CelebA-HQ 256 and FFHQ 256 | FID: 3.54 / 3.66 (ablated) vs 3.72 (full) on CA-L; 3.30 / 3.45 vs 3.65 on FF-L | No clear effect |
| add spectral regularization on top of contrastive loss (ablation) | LC-GAN · FFHQ 256 latent control | number of dimensions to edit per attribute (qualitative) | Improved |
| add contrastive feature categorization into geometry and appearance codes (ablation) | LC-GAN · FFHQ 256 latent control | disentanglement (qualitative) | Improved |
| use LC-GAN latent control instead of SeFa and EigenGAN for viewpoint change | LC-GAN · CelebA-HQ 1024 and FFHQ 1024 | Arcface distance: 0.71 vs 0.91 (SeFa), 0.99 (EigenGAN) at CA-L; 0.61 vs 1.13, 0.99 at FF-H | Improved |
| use LC-GAN instead of SeFa and EigenGAN for appearance change | LC-GAN · CelebA-HQ 1024 | Landmark distance: 5.83 vs 10.42 (SeFa), 13.46 (EigenGAN) | Improved |
| LC-GAN versus uncontrollable StyleNAT at 256 resolution FFHQ | LC-GAN · FFHQ 256x256 | FID: 3.65 vs 2.05 (StyleNAT) | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- StyleSwin — StyleSwin: Transformer-Based GAN for High-Resolution Image Generation (CVPR 2022) · better on 1 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Cluster-Guided Image Synthesis With Unconditional Models (CVPR 2022 · similarity 0.89)
- On Computation and Generalization of Generative Adversarial Networks under Spectrum Control (ICLR 2019 · similarity 0.88)
- Exploring Gradient-Based Multi-directional Controls in GANs (ECCV 2022 · similarity 0.88)
- Improving GAN with Neighbors Embedding and Gradient Matching (AAAI 2019 · similarity 0.88)
- Improved Training of Generative Adversarial Networks Using Representative Features (ICML 2018 · similarity 0.88)
- Large Scale Adversarial Representation Learning (NeurIPS 2019 · 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.
