How ZoomAI Uses AI to Super‑Resolve and Enhance Low‑Quality Videos

Background

Video quality has become a standard expectation for viewers, yet many low‑quality videos and images still exist due to user‑generated content, legacy media conversion, and bandwidth‑saving modes. These issues manifest as low resolution, noise, compression artifacts, dull colors, and scratches.

What Is ZoomAI?

ZoomAI is a comprehensive AI‑based video enhancement suite that applies deep‑learning models to improve image and video quality. Its architecture consists of independent modules—each dedicated to a specific enhancement task such as super‑resolution, noise removal, sharpening, or color enhancement—allowing flexible composition for different scenarios.

Main Modules

Super‑Resolution Module

The module upscales low‑resolution frames to higher resolutions. Training uses an L2 loss to minimize pixel‑wise differences, but L2 alone produces blurry results. ZoomAI mitigates this by (1) adding controlled blur to training data and (2) incorporating a gradient loss that forces the generated image’s gradients to match those of the ground‑truth, yielding sharper outputs.

Denoising and Sharpening Module

Video noise originates from sensor noise, film digitization, and compression artifacts. The module treats denoising as a low‑pass filter while simultaneously training a sharpening branch to preserve or even accentuate edges, preventing the typical loss of detail associated with pure denoising.

Color Enhancement Module

This module addresses under‑exposed, over‑exposed, or dull footage. Two approaches are used: a “black‑box” end‑to‑end network that directly outputs enhanced frames, and a “white‑box” network that predicts explicit parameters (exposure, contrast, saturation) which are then applied to the image. The white‑box design yields smoother color transitions across video frames.

Scratch Removal Module

For archival film restoration, the module leverages temporal information: scratches usually appear in a single frame, so adjacent frames help identify and inpaint missing pixels, effectively removing visible scratches.

Mobile Real‑Time SDK

ZoomAI also provides a lightweight, single‑layer model packaged as a cross‑platform SDK written in OpenGL. The SDK exploits mobile GPU acceleration to perform real‑time 1080p super‑resolution on high‑end iOS and Android devices.

Application Scenarios

Old video restoration: sequentially applying scratch removal, denoising + sharpening, color enhancement, and super‑resolution dramatically reduces manual labor and cost.

Animation upscaling: combining color enhancement with super‑resolution produces high‑definition, vibrant cartoons.

Short‑video cover enhancement: applying denoising + sharpening and color enhancement improves thumbnail quality.

Manga/comic and light‑novel cover upscaling: similar pipelines boost visual appeal.

Future Research Directions

Design additional lightweight modules optimized for mobile deployment.

Develop targeted processing for subtitles, faces, and background blur.

Build an automatic colorization system for black‑and‑white media.

Visual Illustrations