Research map / Image generation / Editing and style transfer
Identity-Preserving Face Recovery from Stylized Portraits
Fatemeh Shiri, Xin Yu, Fatih Porikli, Richard Hartley, Piotr Koniusz
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 discriminative loss to pixel-wise loss (ablation) | IFRP · Seen styles (CelebA test) | FSIM: 0.72 -> 0.75 | Improved |
| add discriminative loss to pixel-wise loss (ablation) | IFRP · Stylized faces, unseen styles (CelebA test) | SSIM: 0.54 -> 0.58 | Improved |
| add identity-preserving loss to pixel and discriminative losses (ablation) | IFRP · Stylized faces, unseen styles (CelebA test) | SSIM: 0.58 -> 0.68 | Improved |
| add identity-preserving loss to pixel and discriminative losses (ablation) | IFRP · Seen styles (CelebA test) | SSIM: 0.62 -> 0.72 | Improved |
| use residual blocks on top-2-layer skip connections instead of a standard autoencoder (ablation) | IFRP · Seen styles (CelebA test) | SSIM: 0.65 -> 0.72 | Improved |
| use residual top-2 skip connections instead of U-net skip connections (ablation) | IFRP · Stylized faces, unseen styles (CelebA test) | SSIM: 0.61 -> 0.68 | Improved |
| add STN layers (ablation) | IFRP (3 training styles) · training convergence | epochs to converge: 165 -> 143 | Improved |
| increase number of training styles from 1 to 3 | IFRP · Stylized faces, unseen styles (CelebA test) | SSIM: 0.54 -> 0.68 | Improved |
| increase number of training styles from 3 to 5 | IFRP · Stylized faces, unseen styles (CelebA test) | SSIM: 0.68 -> 0.69 | No clear effect |
| increase training styles from 1 to 5 raises training time per epoch | IFRP · training time | time per epoch: 1:49' -> 9:47' | Hurt |
| use fewer rotation augmentations (5 angles instead of 11) with STNs (ablation) | IFRP · augmentation study | SSIM: 0.66 -> 0.65 | No clear effect |
Closest papers by meaning
- InverseFaceNet: Deep Monocular Inverse Face Rendering (CVPR 2018 · similarity 0.88)
- Rethinking Deep Face Restoration (CVPR 2022 · similarity 0.88)
- FaceMe: Robust Blind Face Restoration with Personal Identification (AAAI 2025 · similarity 0.87)
- Joint Face Hallucination and Deblurring via Structure Generation and Detail Enhancement (IJCV 2019 · similarity 0.87)
- Face Restoration via Plug-and-Play 3D Facial Priors (TPAMI 2021 · similarity 0.87)
- OSDFace: One-Step Diffusion Model for Face Restoration (CVPR 2025 · 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.
