Knackpad

AI Photo Upscaler
No Upload Required

Enlarge images 2x or 4x with ESRGAN super-resolution — entirely in your browser. Tile-based processing handles any image size without crashing.

100% in-browser — your photos never leave your device
🖼️
Drop an image here or click to browse
JPG · PNG · WebP · Any size (tiled processing handles large images)
Scale factor
Output format
Use-case preset
Select an image to begin.

Which preset should I use?

Each use-case preset adjusts the patch size and quality defaults for the best result.

Preset Best for Recommended settings
Photo Everyday JPG/PNG photos, portraits, landscapes 4x PNG or JPG 90+
Old scan Scanned prints, film grain, faded colors 4x PNG — preserves recovered detail
Digital art Illustrations, pixel art, icons, sprites 2x PNG — avoids over-smoothing crisp lines
Screenshot Low-res UI screenshots, social crops 2x PNG — text edges stay sharp

How it works

ESRGAN (Enhanced Super-Resolution Generative Adversarial Network) is a deep convolutional neural network trained on millions of image pairs. It learns to reconstruct realistic textures — fur, fabric, faces, text edges — rather than blurring pixels outward the way bicubic resampling does.

Step 1
Model download (once)
The ESRGAN Slim model (~1 MB per scale) is fetched from a public CDN and cached by your browser. After the first use it works fully offline.
Step 2
Tiled inference
Large images are split into overlapping patches (128 px + 2 px padding). Each patch runs through the neural network on your GPU (WebGL) or CPU. Tiles are seamlessly stitched, so even large photos process without memory errors.
Step 3
Canvas output
The upscaled tensor is drawn to an HTML5 Canvas, then exported as PNG (lossless) or JPG at your chosen quality level — entirely inside your browser tab.
Step 4
Privacy by design
Your image data never touches a server. You can disconnect from the internet after the model loads — upscaling continues to work perfectly.

Tip: for fastest results keep the input under 1024×1024 px. Larger images still work with tiled processing but take longer — a 2000 px image may take 1–3 minutes on an average laptop.

Frequently asked questions

Yes, completely. The page loads the AI model files (a one-time ~1 MB download) from a public CDN, but your photo never leaves your device. All processing runs in WebGL and JavaScript inside your browser tab. To verify: open DevTools → Network, select an image, click Upscale — you will see zero image-related network requests during processing.
Standard bicubic resampling simply blends neighboring pixels, producing a blurry result. ESRGAN uses a deep neural network trained on millions of high-resolution image pairs. It learns what textures and edges should look like at higher resolution, then reconstructs those details rather than guessing. The result is sharper edges, more realistic texture, and recovered fine detail — especially visible in hair, fabric, and text.
There is no hard limit. This tool uses tiled processing: the image is split into 128 px patches with 2 px overlap, each processed independently, then stitched back together. This means even a 4000 px photo can be upscaled without running out of GPU memory. Processing time scales linearly with pixel count — a 512×512 image takes a few seconds; a 2000 px image may take 1–3 minutes on a typical laptop.
Choose 2x for digital art, screenshots, or when you want speed and sharp crisp lines. Choose 4x for photographs, old scans, or any image where maximum detail recovery matters — such as enlarging small thumbnails or old printed photos. The 4x model takes roughly four times as long because the output has 16× the pixel count. Use the "Digital art" or "Screenshot" presets which default to 2x.
Choose PNG when quality must be preserved exactly — text, graphics, art with sharp edges, or any image you plan to edit further. Choose JPG when file size matters and mild compression is acceptable, such as uploading to social media or a website. The JPG quality slider lets you balance quality vs. file size; 85–92 is near-lossless for most photographs.
Yes. The AI models are cached by your browser after the first download. Once cached, you can disconnect from the internet and upscaling continues to work indefinitely. Each scale (2x and 4x) has its own model; both are cached separately on first use.