Research map / 3D vision / Face and shape
3D-FM GAN: Towards 3D-Controllable Face Manipulation
Yuchen Liu, Zhixin Shu, Yijun Li, Zhe Lin, Richard Zhang, S.Y. Kung
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 |
|---|---|---|---|
| encode photo into W+ instead of render into W (ablation) | exclusive modulation · FFHQ held-out 5k, manipulation | identity preservation (Id): 0.57 -> 0.66 | Improved |
| encode photo into W+ instead of render into W (ablation) | exclusive modulation · FFHQ held-out 5k, manipulation | landmark error (LM): 17.8 -> 27.3 | Hurt |
| encode render into W+ instead of W (ablation) | exclusive modulation · FFHQ held-out 5k, manipulation | FID: 14.5 -> 25.8 | Hurt |
| use multiplicative instead of concatenation co-modulation (ablation) | 2-encoder co-modulation · FFHQ held-out 5k, manipulation | FID: 17.8 -> 12.4 | Improved |
| use multiplicative instead of concatenation co-modulation (ablation) | 2-encoder co-modulation · FFHQ held-out 5k, manipulation | landmark error (LM): 59.3 -> 22.2 | Improved |
| use multiplicative instead of tensor transform co-modulation (ablation) | 2-encoder co-modulation · FFHQ held-out 5k, manipulation | FID: 18.6 -> 12.4 | Improved |
| add third encoder to multiplicative co-modulation (ablation) | co-modulation · FFHQ held-out 5k, manipulation | landmark error (LM): 22.2 -> 17.2 | Improved |
| add third encoder to multiplicative co-modulation (ablation) | co-modulation · FFHQ held-out 5k, manipulation | face content similarity error (FC): 0.025 -> 0.020 | Improved |
| add third encoder to multiplicative co-modulation (ablation) | co-modulation · FFHQ held-out 5k, manipulation | identity preservation (Id): 0.66 -> 0.66 | No clear effect |
| alternate reconstruction and disentanglement training instead of either alone (ablation) | Render-W · synthetic data, 140k iterations | identity and editability | Improved |
| add real-data (FFHQ) reconstruction to training (ablation) | Render-W · 280k iterations | photo-realism | Improved |
| two-phase (synthetic then real) instead of real-only reconstruction (ablation) | Render-W · 280k iterations | identity preservation | Improved |
Closest papers by meaning
- MOST-GAN: 3D Morphable StyleGAN for Disentangled Face Image Manipulation (AAAI 2022 · similarity 0.91)
- Injecting 3D Perception of Controllable NeRF-GAN into StyleGAN for Editable Portrait Image Synthesis (ECCV 2022 · similarity 0.90)
- Mask-Guided Portrait Editing With Conditional GANs (CVPR 2019 · similarity 0.90)
- In-N-Out: Faithful 3D GAN Inversion with Volumetric Decomposition for Face Editing (CVPR 2024 · similarity 0.90)
- CGS-GAN: 3D Consistent Gaussian Splatting GANs for High Resolution Human Head Synthesis (NeurIPS 2025 · similarity 0.88)
- Reference-Based 3D-Aware Image Editing with Triplanes (CVPR 2025 · 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.
