Research map / Image generation / Editing and style transfer
One-Shot Neural Face Reenactment via Finding Directions in GAN’s Latent Space
Stella Bounareli, Christos Tzelepis, Vasileios Argyriou, Ioannis Patras, Georgios Tzimiropoulos
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 generator weight optimization at inference (ablation) | LOR · VoxCeleb1 self-reenactment | CSIM: 0.37 -> 0.66 | Improved |
| add joint training of inversion encoder and directions (no optimization) (ablation) | LOR+ · VoxCeleb1 self-reenactment | ARD: 1.1 -> 0.8 | Improved |
| add feature space refinement (FSR) (ablation) | LOR+ · VoxCeleb1 self-reenactment | ARD: 0.8 -> 0.7 | Improved |
| add feature space refinement (FSR) (ablation) | LOR+ · VoxCeleb1 self-reenactment, large head pose (Benchmark-L) | CSIM: 0.60 -> 0.62 | Improved |
| add identity and perceptual losses (ablation) | directions pipeline · Synthetic image pairs (5K), cross-subject (ablation) | CSIM: 0.42 -> 0.52 | Improved |
| add eye and mouth losses (ablation) | directions pipeline · Synthetic image pairs (5K), cross-subject (ablation) | AED: 1.5 -> 1.2 | Improved |
| add real images to synthetic training (ablation) | directions pipeline · VoxCeleb1 self-reenactment (training data ablation) | CSIM: 0.60 -> 0.63 | Improved |
| fine-tune on paired video data (ablation) | directions pipeline · VoxCeleb1 self-reenactment (training data ablation) | ARD: 1.6 -> 1.1 | Improved |
| add style loss (ablation) | LOR+ joint training · VoxCeleb1 self-reenactment, joint training (loss ablation) | CSIM: 0.64 -> 0.66 | Improved |
| add cycle loss (ablation) | LOR+ joint training · VoxCeleb1 self-reenactment, joint training (loss ablation) | CSIM: 0.62 -> 0.66 | Improved |
| add single-attribute change in 50% of training samples (ablation) | directions matrix · Disentanglement (synthetic, yaw changed) | pitch error: 1.3 -> 1.2 | Improved |
| use LOR+ instead of LOR (joint training + FSR, no optimization) | reenactment · VoxCeleb1 cross-subject reenactment | FID: 86.2 -> 78.4 | Improved |
Closest papers by meaning
- HyperReenact: One-Shot Reenactment via Jointly Learning to Refine and Retarget Faces (ICCV 2023 · similarity 0.91)
- InterFaceGAN: Interpreting the Disentangled Face Representation Learned by GANs (TPAMI 2020 · similarity 0.91)
- FSGANv2: Improved Subject Agnostic Face Swapping and Reenactment (TPAMI 2022 · similarity 0.90)
- FSGAN: Subject Agnostic Face Swapping and Reenactment (ICCV 2019 · similarity 0.89)
- VOODOO 3D: Volumetric Portrait Disentanglement For One-Shot 3D Head Reenactment (CVPR 2024 · similarity 0.89)
- In-N-Out: Faithful 3D GAN Inversion with Volumetric Decomposition for Face Editing (CVPR 2024 · similarity 0.87)
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.
