Computer Vision / Project
Nostalix · Image & video restoration
Image and video restoration with RRDBNet, Real-ESRGAN and GFPGAN, served through a Flask API.
- My role
- Machine learning & application development
- Period
- Dec 2025 — Jan 2026
- PSNR
- ≈28.5 dB
- SSIM
- ≈0.92
- Scale factor
- 4×
Project evaluation figures. 4× denotes the output scale factor; restoration quality depends on the input data.
01 / The context
What the project addresses
A restoration application for degraded images and videos, including historical archives, general imagery and anime / manga.
02 / My contribution
What I worked on
Prepared image data, trained and evaluated restoration models, and integrated pretrained models into an API and web application.
The challenge
Connect restoration quality with practical image / video inference and available GPU memory.
03 / The approach
How the pieces fit together
- Prepare a 52,000-image FFHQ dataset with normalization, resizing, patch extraction and augmentation.
- Train RRDBNet models with PyTorch and evaluate PSNR, SSIM and MSE.
- Serve four pretrained Real-ESRGAN / GFPGAN models through a Flask REST API.
- Support image / video batches, automatic GPU detection, adaptive memory handling and 4× scaling.
04 / The outcome
What the work delivered
- MSE of approximately 0.012 in the reported evaluation.
- A web application connecting restoration models to batch image and video processing.
View the Nostalix application screenshot
Next case study