Super-Resolution & Restoration
Make small, blurry or old images sharp
WHAT THIS CAPABILITY MEANS
Takes a low-resolution image or a degraded one — blurry, noisy, scratched — and outputs a higher-resolution or cleaner version. It is not mere upscaling: it fills high-frequency detail from learned image priors, and precisely because it fills, some of it is guesswork. Unlike image-to-image the goal is restoration, not stylisation.
How it is done
The classic route regresses high-resolution pixels with convolutional networks and leans on perceptual and adversarial losses for crisper texture, trading pixel fidelity for realism. Diffusion models were later adopted to fill detail from a generative prior, doing better on severe degradation. Because real-world degradation is complex and unknown, training often randomises the degradation so the model does not depend on one fixed downsampling.
Representative products
4Stable Diffusion
2022Released text-to-image weights openly and small enough to run on consumer GPUs
FLUX
2024Generates high-resolution images with a rectified-flow transformer
Diffusers
2022A unified implementation and scheduler for diffusion models
Firefly
2023An image generation and editing tool aimed at creators
Organizations involved
Typical uses
- Restoring old photos and family footage
- Detail enhancement for surveillance and satellite imagery
- Sharpening medical and microscopy images
- Upscaling assets for print and publishing
How it is evaluated
- PSNR
- Peak signal-to-noise ratio, a pixel-level fidelity measure
- SSIM
- Structural similarity, closer to human judgement than PSNR
- LPIPS
- Perceptual distance, judging naturalness of texture
Limits and hard parts
- It invents texture that was never there; enlarged faces and text can become something else entirely
- High PSNR does not mean it looks good; over-smoothed outputs score well yet lack detail
- Generalisation is poor on degradation types unseen in training, such as a particular compression artefact
Concepts behind it
Image Representation
To a machine, a photo is nothing but stacked grids of numbers
Convolutional Neural Networks
Replacing full connections with “look locally, reuse the same filter everywhere” — the idea that made image recognition work
Generative Adversarial Networks
A forger versus an inspector: each pushes the other to its limit