Research map / Image generation / GANs
Exploring Gradient-Based Multi-directional Controls in GANs
Zikun Chen, Ruowei Jiang, Brendan Duke, Han Zhao, Parham Aarabi
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 gradient-based channel exclusion (top-k salient dimensions of entangled attributes) (ablation) | gradient direction for gender · FFHQ StyleGAN2 (gender edit) | entanglement with smile, eyeglasses, age (qualitative) | Improved |
| exclude top-k channels with k set to 100 for all attributes (ablation) | gradient direction · FFHQ StyleGAN2 | disentanglement (qualitative) | Improved |
| tune k per attribute in {50,100,150,200,250} on 30-example meta-validation set (ablation) | channel exclusion · FFHQ StyleGAN2 | disentanglement (qualitative) | Improved |
| use gradient directions instead of InterFaceGAN SVM directions, gender | InterFaceGAN · FFHQ StyleGAN2 gender | manipulation accuracy: 0.5859 -> 0.6937 | Improved |
| use gradient directions instead of InterFaceGAN, smile | InterFaceGAN · FFHQ StyleGAN2 smile | manipulation accuracy: 0.9033 -> 0.9254 | Improved |
| use gradient directions instead of InterFaceGAN, eyeglasses | InterFaceGAN · FFHQ StyleGAN2 eyeglasses | manipulation accuracy: 0.5946 -> 0.6626 | Improved |
| use gradient directions instead of InterFaceGAN, age | InterFaceGAN · FFHQ StyleGAN2 age | manipulation accuracy: 0.5620 -> 0.6139 | Improved |
| use gradient directions instead of StyleSpace, age | StyleSpace · FFHQ StyleGAN2 age | manipulation accuracy: 0.0034 -> 0.6139 | Improved |
| use gradient directions instead of StyleSpace, eyeglasses | StyleSpace · FFHQ StyleGAN2 eyeglasses | attribute dependency (mean AD) | Hurt |
| use gradient directions instead of InterFaceGAN for gender and age as step grows | InterFaceGAN · FFHQ StyleGAN2 | attribute dependency (mean AD) | Improved |
| use gradient directions instead of StyleSpace/GANSpace for smile | StyleSpace and GANSpace · FFHQ StyleGAN2 smile | attribute dependency (mean AD) | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- InterFaceGAN — InterFaceGAN: Interpreting the Disentangled Face Representation Learned by GANs (TPAMI 2020) · better on 4 of 4 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
- LARGE: Latent-Based Regression Through GAN Semantics (CVPR 2022 · similarity 0.90)
- InfoGAN-CR and ModelCentrality: Self-supervised Model Training and Selection for Disentangling GANs (ICML 2020 · similarity 0.90)
- GANSpace: Discovering Interpretable GAN Controls (NeurIPS 2020 · similarity 0.89)
- On the Statistical Properties of Generative Adversarial Models for Low Intrinsic Data Dimension (JMLR 2025 · similarity 0.89)
- Interpreting the Latent Space of GANs for Semantic Face Editing (CVPR 2020 · similarity 0.89)
- A Classification-Based Study of Covariate Shift in GAN Distributions (ICML 2018 · 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.
