FaceMe: Robust Blind Face Restoration with Personal Identification
Authors: Siyu Liu, Zheng-Peng Duan, Jia OuYang, Jiayi Fu, Hyunhee Park, Zikun Liu, Chun-Le Guo, Chongyi Li
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our Face Me can restore high-quality facial images while maintaining identity consistency, achieving excellent performance and robustness. ... We compare Face Me with state-of-the-art methods, including GFP-GAN (Wang et al. 2021), Code Former (Zhou et al. 2022), VQFR (Gu et al. 2022), GPEN (Yang et al. 2021), Dif Face (Yue and Loy 2024), and DR2 (Wang et al. 2023), PGDiff (Yang et al. 2024), and DMDNet (Li et al. 2022). ... Ablation Studies |
| Researcher Affiliation | Collaboration | Siyu Liu1*, Zheng-Peng Duan1*, Jia Ou Yang2, Jiayi Fu1, Hyunhee Park3, Zikun Liu2, Chun-Le Guo1, 4 , Chongyi Li1, 4 1VCIP, CS, Nankai University 2Samsung Research, China, Beijing (SRC-B) 3The Department of Camera Innovation Group, Samsung Electronics 4NKIARI, Shenzhen Futian EMAIL, EMAIL, EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes methods and training strategies in text and diagrams (Figure 2), but does not contain a clearly labeled pseudocode or algorithm block. |
| Open Source Code | No | Code https://modyu-liu.github.io/Face Me Homepage. This URL points to a project homepage, which is considered a project demonstration page or high-level overview page rather than a direct link to a code repository according to the provided instructions. |
| Open Datasets | Yes | Our training dataset consists of FFHQ dataset (Karras, Laine, and Aila 2019) and our synthesized FFHQRef dataset... Testing datasets We use one synthetic dataset Celeb Ref HQ (Li et al. 2022) and three real-world datasets: LFW-Test (Huang et al. 2008), Web Photo-Test (Wang et al. 2021), and WIDER-Test (Zhou et al. 2022) for test. Celeb Ref-HQ is collected by crawling images of celebrities from the internet. It contains 1,005 identities and a total of 10,555 images. LFW-Test consists of 1,711 mildly degraded face images from the LFW dataset. Web Photo-Test consists of 407 medium degraded face images from the internet. WIDER-Test consists of 970 severely degraded face images from the WIDER Face (Yang et al. 2016) dataset. |
| Dataset Splits | Yes | Our training dataset consists of FFHQ dataset (Karras, Laine, and Aila 2019) and our synthesized FFHQRef dataset, with all images resized to 512 512. Ih denotes the high-quality image from the FFHQ dataset. To form training pairs, 1 4 images with the same identity as Ih are randomly selected from the FFHQRef dataset as reference images. ... For the synthetic dataset, we randomly select 150 identities and select one image per identity as the ground truth, using 1 4 images of the same identity as reference images. |
| Hardware Specification | Yes | The training process is implemented using the Py Torch framework and is conducted on eight A40 GPUs, with a batch size of 4 per GPU. |
| Software Dependencies | No | We employ the SDXL model (Podell et al. 2023) stable-diffusion-xl-base-1.0 fine-tuned by Photo Maker as our base diffusion model. We employ the CLIP image encoder, fine-tuned by Photo Maker, as part of our identity encoder. We use the Adam W (Loshchilov and Hutter 2019) optimizer to optimize the network parameters... The training process is implemented using the Py Torch framework... While specific models like SDXL base-1.0 are mentioned, general software dependencies like PyTorch are not provided with version numbers. |
| Experiment Setup | Yes | We use the Adam W (Loshchilov and Hutter 2019) optimizer to optimize the network parameters with a learning rate of 5 10 5 for two training stages. ... The two training stages are trained 130K and 210K iterations, respectively. ... with a batch size of 4 per GPU. |