Image Upscaler
Enlarge an image 2× to 8× with Lanczos-3, Catmull-Rom bicubic or nearest-neighbour resampling, with optional sharpening and a zoomed comparison. This is classic interpolation, not AI: edges stay clean and sharp, but no new detail is invented.
How to Use
- Drop an image on the box, click to browse, or paste one. Any format the site’s image engine reads works, including HEIC, TIFF and PSD. A built-in sample is enlarged first.
- Choose the Factor (2× to 8×) and the Method: Lanczos-3 for photos, bicubic for a slightly softer result, nearest neighbour for pixel art and screenshots of pixels.
- Set Sharpen if the result looks soft; 20–40% suits most photos. Leave it at 0 for pixel art.
- Click the result to choose a spot, and compare it at 2× zoom against another method.
- Pick an output format and download. The output is limited to 32 megapixels so the browser tab does not run out of memory.
Worked Example
The 640 × 420 sample at 2× with Lanczos-3. The output is 1280 × 840 = 1,075,200 pixels (1.08 MP, 4.1 MB as RGBA). Output column 11 sits at source position (11 + 0.5) ÷ 2 − 0.5 = 5.25, so its colour is 0.8928 × pixel 5 + 0.2710 × pixel 6 − 0.1333 × pixel 4 − 0.0680 × pixel 7 + 0.0301 × pixel 3 + 0.0074 × pixel 8. The weights come from L(x) = sinc(x) · sinc(x ÷ 3) and are divided by their sum (0.997) so a flat colour stays exactly the same. The two small negative weights are what keep edges crisp.
The 16 × 16 pixel-art heart at 8× with nearest neighbour. The output is 128 × 128. Each output pixel takes the colour of source pixel ⌊(x + 0.5) ÷ 8⌋, so every source pixel becomes an exact 8 × 8 block, and the picture still uses only its original three colours plus transparency.
The common mistake: smoothing pixel art. The same heart enlarged 8× with Lanczos-3 gets soft edges, in-between shades and faint rings next to the black outline, because the filter is designed for continuous photos. The "Pixel art the wrong way" preset shows it. And no method recovers detail: a 640 px photo enlarged to 2560 px has 16 times the pixels but not more information.
Show Work
Formulas
Interpolation, Not Invention
Enlarging a picture means estimating colours between the pixels you have. The ideal filter for a band-limited signal is the sinc function, but it extends forever, so practical filters cut it off. Lanczos resampling multiplies sinc by a wider sinc as a window, a technique named after the mathematician Cornelius Lanczos; with a window of 3 it looks at 3 pixels on each side. Bicubic convolution as used here follows Robert Keys' 1981 paper on cubic convolution for image processing; with a = −0.5 it is the Catmull-Rom spline from 1974. Nearest neighbour does no estimating at all.
All three are deterministic: the same input always gives the same output, and each output pixel depends only on a few nearby input pixels. AI upscalers work differently: a neural network trained on many photos predicts what fine detail probably looked like. That can be impressive on faces and textures, but the added detail is a guess, which matters for documents, evidence and anything that has to stay faithful to the original.
About This Tool
This upscaler enlarges an image 2× to 8× with its own separable Lanczos-3, Catmull-Rom bicubic or nearest-neighbour resampler, run in a background worker on premultiplied alpha, with an optional unsharp mask. Show Work lists the real filter weights for the first output columns, and a zoomed crop compares the result with another method.
Images are decoded and saved in your browser by the engine behind the Universal Image Converter, so HEIC, TIFF, PSD and RAW files work. Nothing is uploaded. Settings can go in the page address, for example ?factor=4&method=nearest&sharp=0.
Related tools: Image Resizer, Pixel Art Editor, and Image Adjustments.
Frequently Asked Questions
Is this AI upscaling?
No. Every new pixel is a weighted average of nearby source pixels, using a fixed mathematical filter. The result is clean and sharp-edged but contains no detail that was not in the original. AI super-resolution models guess plausible texture, hair or text, and sometimes guess wrong; this tool never adds anything that was not there.
What is Lanczos-3 and why is it the default?
Lanczos-3 uses the kernel L(x) = sinc(x) · sinc(x ÷ 3) over 3 source pixels on each side, so 6 to 7 taps per axis. Its small negative lobes keep edges crisp, so it looks sharper than bicubic. The price is a faint halo next to very hard edges, which is why pixel art should use nearest neighbour instead.
Which method is right for pixel art?
Nearest neighbour at a whole-number factor. Each source pixel becomes an exact square block, 8 × 8 at 8×, with no new colours. Lanczos or bicubic would blur the edges between colours and add in-between shades that pixel art does not have.
How large can the output be?
Up to 32 megapixels, for example 6,000 × 5,333 px. At 4 bytes per pixel that is about 128 MB of image data before sharpening, which most phones and computers handle. A 4000 × 3000 photo (12 MP) therefore cannot be doubled here, because 8000 × 6000 is 48 MP; the Image Resizer can make smaller enlargements such as 1.5×.
Is transparency kept?
Yes. The resampler works on premultiplied alpha: colours are weighted by their opacity before averaging, so a red shape on a transparent background gets clean red edges instead of a dark fringe. Save as PNG, WebP or another format with alpha to keep it.
How do I use the Image Upscaler?
Just type your numbers. The answer shows up right away — there is no button to press. Change anything and it updates by itself.
Do I need to install or sign up for anything?
Not at all — it runs in the browser with nothing to install and no account. After it loads once, it even works without an internet connection.
Is my information private?
Yes. Everything happens in your browser. Nothing you type is sent to a server or saved anywhere.
Common Use Cases
Pixel-art sprites
A 16 × 16 sprite at 8× with nearest neighbour becomes 128 × 128 with every pixel an exact 8 × 8 square.
Small web images for print
A 1200 × 800 image at 3× becomes 3600 × 2400, enough pixels for a 12 × 8 in print at 300 dpi, though smooth rather than more detailed.
High-DPI assets
Make a 2× version of a 240 × 80 logo (480 × 160) when the original vector file is lost.
Comparing methods
Zoom into the same 128 px crop to see nearest, bicubic and Lanczos-3 side by side on your own picture.
Last updated: