Research map / Image generation / GANs
Rewriting a Deep Generative Model
David Bau, Steven Liu, Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba
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 |
|---|---|---|---|
| optimize the weight change directly (direct optimization of Lambda) instead of the constrained projected gradient descent (ablation) | rank-one model rewriting · FFHQ smile edit | LPIPS (masked): 0.035084010987915096 -> 0.1375907236509025 | Hurt |
| add the single-layer-without-direction-constraint variant of the edit (remove the direction constraint) (ablation) | rank-one model rewriting · LSUN church dome to spire | LPIPS (masked): 0.020296541112358683 -> 0.294 | Hurt |
| constrain the update to the user-specified context direction rather than a single-image direction (ablation) | rank-one model rewriting · LSUN church dome to spire | % dome pixels correctly modified: 90.14 -> 92.03 | Improved |
| use constrained rank-one update instead of finetuning all weights | StyleGANv2 FFHQ · FFHQ smile edit | LPIPS (masked): 0.40296847543120384 -> 0.035084010987915096 | Improved |
| use our method instead of NBB with Laplace blending | StyleGANv2 LSUN church · LSUN church dome to tree | % dome pixels correctly modified: 31.18230837 -> 48.65 | Improved |
| use our method instead of NBB for smile edits | StyleGANv2 FFHQ · FFHQ smile edit | % smiling images: 93.51 -> 84.37 | Hurt |
| replace direct optimization of Lambda with projected gradient descent (ablation) | rank-one model rewriting · LSUN church dome to tree | % dome pixels correctly modified: 59.42852977622276 -> 48.65 | Mixed |
| apply rank-1 update instead of zeroing 30 GAN Dissection units for watermark removal | StyleGANv2 LSUN church · Watermark removal (bottom text) | count of visible watermarks: 6 -> 0 | Improved |
| apply rank-1 update instead of zeroing 60 GAN Dissection units for watermark removal | StyleGANv2 LSUN church · Watermark removal (bottom text) | count of visible watermarks: 4 -> 0 | Improved |
Closest papers by meaning
- Linear Semantics in Generative Adversarial Networks (CVPR 2021 · similarity 0.87)
- Navigating the GAN Parameter Space for Semantic Image Editing (CVPR 2021 · similarity 0.85)
- Transferring GANs: generating images from limited data (ECCV 2018 · similarity 0.85)
- GLeaD: Improving GANs With a Generator-Leading Task (CVPR 2023 · similarity 0.85)
- Controlling generative models with continuous factors of variations (ICLR 2020 · similarity 0.85)
- Minimalist Concept Erasure in Generative Models (ICML 2025 · similarity 0.84)
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.
