Abstract
Nowadays, due to the ubiquitous visual media there are vast amounts of already available high-resolution (HR) face images. Therefore, for super-resolving a given very low-resolution (LR) face image of a person it is very likely to find another HR face image of the same person which can be used to guide the process. In this paper, we propose a convolutional neural network (CNN)-based solution, namely GWAInet, which applies super-resolution (SR) by a factor 8x on face images guided by another unconstrained HR face image of the same person with possible differences in age, expression, pose or size. GWAInet is trained in an adversarial generative manner to produce the desired high quality perceptual image results. The utilization of the HR guiding image is realized via the use of a warper subnetwork that aligns its contents to the input image and the use of a feature fusion chain for the extracted features from the warped guiding image and the input image. In training, the identity loss further helps in preserving the identity related features by minimizing the distance between the embedding vectors of SR and HR ground truth images. Contrary to the current state-of-the-art in face super-resolution, our method does not require facial landmark points for its training, which helps its robustness and allows it to produce fine details also for the surrounding face region in a uniform manner. Our method GWAInet produces photo-realistic images in upscaling factor 8x and outperforms state-of-the-art in quantitative terms and perceptual quality.
Introduction
Face "hallucination"—the process of reconstructing high-frequency details from very low-resolution (LR) images—is a massive challenge in computer vision. Standard super-resolution (SR) techniques often produce "smooth" or blurry results because they optimize for pixel-level similarity (like PSNR) rather than perceptual sharpness. While using a reference (exemplar) photo of the same person can help guide the reconstruction of identity-specific details, existing methods usually depend on facial landmarks (eyes, nose, mouth points) to align the reference image to the blurry input.
This creates a significant bottleneck: landmark detectors perform poorly on the very low-resolution images that need super-resolution the most. Furthermore, landmark-based methods often only sharpen the areas around those specific points, leaving the rest of the face looking inconsistent. Researchers from ETH Zurich have developed GWAInet, a deep learning model that performs 8x super-resolution on tiny, blurry face images by using a high-resolution guiding photo without relying on landmark detection.
Methodology
The authors propose GWAInet, an end-to-end architecture composed of four specialized subnetworks designed to handle alignment, feature fusion, and identity preservation.
- Warper (Wnet): Instead of using landmarks, this subnetwork learns a "flow field" to spatially align the high-resolution (HR) guiding image to the LR input. It accounts for differences in pose, expression, and scale, providing a robust foundation for feature extraction.
- Generator (Gnet): This is the heart of the system. It consists of two branches: SRnet, which processes the LR input, and GFEnet, which extracts features from the warped guiding image. These features are fused at multiple scales to reconstruct the final image, allowing the model to integrate high-frequency details from the guide into the low-resolution structure.
- Critic (Cnet): To ensure photo-realism, the model is trained as a Generative Adversarial Network (GAN) using the WGAN-GP framework, which helps the model learn the distribution of real human faces and generate sharp, natural textures.
- Identity Encoder (Inet): A pretrained Siamese network that calculates an "identity loss." It ensures that the embedding vector (the mathematical "fingerprint") of the generated image matches the ground truth, preserving the person's unique look.
Experiments and Results
The model was developed and evaluated using several large-scale face datasets, including CelebA, VggFace2, and CASIA-WebFace. The authors performed 8x upscaling, taking tiny 32x32 inputs and producing 256x256 outputs.
Quantitative Superiority
GWAInet outperformed the previous state-of-the-art (GFRNet) by a significant margin. On the VggFace2 dataset, it achieved a PSNR of 25.57 dB, compared to GFRNet's 24.10 dB. This improvement demonstrates the effectiveness of the landmark-free warping and multi-scale feature fusion.
Qualitative Sharpness and Identity Preservation
Visually, GWAInet produces significantly sharper results for 8x upscaling. While other methods often produce artifacts or "partially sharp" faces, GWAInet generates consistent textures for hair, wrinkles, and skin. By ditching landmark detectors in favor of a learned warping network, the model avoids the "brittleness" of traditional alignment. This allows the network to produce uniform, sharp details across the entire face, including hair and skin texture, rather than just around the eyes or mouth. Furthermore, the inclusion of the identity loss forces the GAN to prioritize "looking like the right person" over just "looking like a generic sharp face."
Conclusion
GWAInet represents a significant step forward in exemplar-guided face super-resolution. By removing the dependency on facial landmarks, the model achieves greater robustness and produces more consistent, photo-realistic results. While the model currently requires an exemplar image of the same person, its ability to handle extreme upscaling factors while preserving identity makes it a powerful tool for facial reconstruction tasks. Future work may explore synthesizing details when a specific exemplar is missing or further improving the fusion of features for even higher upscaling factors.
Summary updated