Face Super-Resolution Demo

Upload a face image to enhance its resolution using our trained deep learning model. Compare results against traditional upscaling methods (Bicubic, Lanczos).

How to use:

Upload an image OR select a sample from the validation set
Choose a model (if multiple are available)
Click "Enhance" to process

Note: For accurate metrics, upload a 256x256 image (will be downscaled to 64x64, then upscaled back)

Input

Upload Face Image

Or select a sample

Select Model

Show Bicubic

Show Lanczos

Results

Low Resolution Input

Super-Resolution Output

Comparison (scroll to see all)

Metrics

About the Metrics

Metric	Description	Better
PSNR	Peak Signal-to-Noise Ratio - measures pixel-level accuracy	Higher
SSIM	Structural Similarity - measures structural preservation	Higher
LPIPS	Learned Perceptual Image Patch Similarity - measures perceptual quality	Lower

Note: GAN-trained models often have slightly lower PSNR/SSIM but much better LPIPS, because they generate sharper, more realistic details that may not exactly match the ground truth pixels.