Research map / 3D vision / Face and shape
Normalized Avatar Synthesis Using StyleGAN and Perceptual Refinement
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use the proposed StyleGAN2-based normalized avatar method instead of Deng et al. | 3D face reconstruction · FaceScape | average point-to-mesh distance (mm): 1.568mm -> 1.557mm | Helped |
| Use the proposed StyleGAN2-based normalized avatar method instead of Tran et al. | 3D face reconstruction · FaceScape | average point-to-mesh distance (mm): 1.935mm -> 1.557mm | Helped |
| Use the proposed normalized texture method instead of Deng et al. | Texture reconstruction · FaceScape | mean L1 pixel loss: 0.392 -> 0.205 | Helped |
| Use the proposed normalized texture method instead of Tran et al. | Texture reconstruction · FaceScape | mean L1 pixel loss: 0.304 -> 0.205 | Helped |
| Add perceptual loss to pixel and adversarial losses during GAN inversion (ablation) | StyleGAN2 latent-code inversion · Rendered views of reconstructed geometry and texture | high-level appearance similarity | Helped |
| Add perceptual refinement after inference (ablation) | Normalized avatar synthesis · Unconstrained input photographs | likeness and expressiveness | Helped |
| Initialize refinement from the identity regressor prediction instead of a mean latent vector (ablation) | Perceptual refinement · Unconstrained input photographs | faithful face reconstruction: 200 iterations | Helped |
| Train synthesis and regression networks with the Normalized Face Dataset instead of only the Scan Dataset (ablation) | StyleGAN2 synthesis network and identity regressor · Normalized face reconstruction | expressiveness and reconstruction accuracy | Helped |
| Use nonlinear inference geometry to initialize Face2Face optimization instead of a linear 3DMM initialization (ablation) | Face2Face optimization · Face geometry fitting | preservation of nonlinear geometry and detail | Helped |
| Use pixel loss and adversarial loss without perceptual loss during GAN inversion (ablation) | StyleGAN2 latent-code inversion · Rendered views of reconstructed geometry and texture | overall similarity | Hurt |
| Use identity loss and latent regularization without perceptual loss during refinement (ablation) | Perceptual refinement · Unconstrained input photographs | likeness and asset quality | Hurt |
| Use perceptual loss and latent regularization without identity loss during refinement (ablation) | Perceptual refinement · Unconstrained input photographs | identity consistency | Hurt |
Closest papers by meaning
- Unsupervised Face Normalization With Extreme Pose and Expression in the Wild (CVPR 2019 · similarity 0.90)
- AniFaceGAN: Animatable 3D-Aware Face Image Generation for Video Avatars (NeurIPS 2022 · similarity 0.89)
- High-Fidelity 3D Head Avatars Reconstruction through Spatially-Varying Expression Conditioned Neural Radiance Field (AAAI 2024 · similarity 0.87)
- Gaussian Eigen Models for Human Heads (CVPR 2025 · similarity 0.87)
- CGS-GAN: 3D Consistent Gaussian Splatting GANs for High Resolution Human Head Synthesis (NeurIPS 2025 · similarity 0.86)
- MoSAR: Monocular Semi-Supervised Model for Avatar Reconstruction using Differentiable Shading (CVPR 2024 · similarity 0.86)
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.