Перейти к содержимому
Атлас ИИ

Суперразрешение и восстановление

Сделать чёткими мелкие, размытые или старые изображения

Генерация и редактирование изображенийНачальный #21
входИзображениеИзображение

Полный текст статьи представлен на английском; заголовок и аннотация локализованы.

ЧТО ЭТО ЗА ВОЗМОЖНОСТЬ

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.

Как это устроено

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.

Примеры продуктов

4

Связанные организации

Типичное применение

  • Restoring old photos and family footage
  • Detail enhancement for surveillance and satellite imagery
  • Sharpening medical and microscopy images
  • Upscaling assets for print and publishing

Как её оценивают

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

Границы и трудности

  • 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

Концепции в основе