How It Works

AI Image Upscaling

ESRGAN deep learning via upscaler.js + WebAssembly — 2× super-resolution running privately in your browser.

ESRGAN Neural Network

Enhanced Super-Resolution GAN — a deep learning model that reconstructs missing detail rather than just stretching pixels.

  • RRDB blocks capture multi-scale features for realistic detail
  • Perceptual loss — optimizes for visual quality not just pixels
  • 2× super resolution with intelligent detail reconstruction
  • Trained on millions of image pairs for photorealistic output

WebAssembly Runtime

Upscaler.js with WASM backend runs ESRGAN directly in the browser — no GPU, no server, no uploads.

  • Client-side AI inference via upscaler.js
  • Patch-based processing handles large images in tiles
  • Seamless merging of processed tiles with no visible seams
  • Cached after first load — runs fully offline

AI upscaling pipeline

1

Image Load

Decode & preprocess

2

Patch Split

Divide into tiles

3

Feature Extract

RRDB analysis

4

Detail Generate

AI upscaling

5

Patch Merge

Seamless blend

6

Export

PNG/WebP/JPEG

Upscaling methods comparison

MethodTechnologyQualitySpeedBest for
BicubicInterpolation★★InstantSimple scaling, no detail needed
LanczosAdvanced interpolation★★★InstantQuick resizing, sharper than bicubic
ESRGAN (imgshrnk)Deep Learning + WASM★★★★★5–60 sHigh-quality enhancement with detail reconstruction
ESRGAN (GPU)Deep Learning + CUDA★★★★★1–10 sFastest AI upscaling (requires powerful GPU)

Technical specifications

Input formatsJPEG, PNG, WebP, AVIF
Output formatsPNG (lossless), WebP, JPEG
Recommended input512–2048 px for best performance
Output size2× input dimensions (4× pixel count)
512×512 processing5–15 s
1024×1024 processing15–30 s
Model download~5 MB (cached after first use)
AI package@upscalerjs/esrgan-thick

Use cases

Photo Enlargement

Prepare photos for large-format printing or professional displays.

Retina Displays

Create sharp graphics for high-DPI screens and responsive web design.

Digital Art

Scale artwork and illustrations for different display sizes.

Product Photos

Enhance e-commerce images for zoom features and detail views.

Social Media

Upscale images for Instagram, LinkedIn, and Twitter with enhanced details.

Video Thumbnails

Create high-quality thumbnails from video frames for YouTube and Vimeo.

ESRGAN, not bicubic

Unlike simple interpolation, ESRGAN reconstructs plausible detail from learned patterns — the difference is visible at high zoom.

No GPU required

ESRGAN runs via WebAssembly on any modern CPU. Slower than a dedicated GPU but private, free, and with zero setup.

~5 MB model

The ESRGAN-thick model is downloaded once and cached. All subsequent upscales are instant to start.