How to upscale a low-resolution image (and when not to bother)

What upscaling can and cannot do

Upscaling a 500x500 image to 2000x2000 quadruples the pixel count. Those extra pixels have to come from somewhere. There are three approaches:

1. Nearest-neighbor / bilinear scaling. Just duplicate or interpolate between existing pixels. Fast, universal, produces blocky or blurry results. This is what happens when you drag an image bigger in most software.

2. Lanczos / bicubic scaling. Fancier interpolation that produces smoother results. Better than bilinear, but still fundamentally limited — sharp lines stay sharp, but detail that wasn't there stays not-there.

3. AI upscaling (super-resolution). Neural networks trained on millions of high/low-res image pairs learn to predict what a higher-resolution version would look like. Results are dramatically better than interpolation on faces, text, and natural scenes. Can produce hallucinations on tricky content (fine hair, unfamiliar text).

Image Upscaler uses the third approach.

When upscaling helps

Realistic wins:

• Recovering usable resolution from small web images. A 400x400 profile picture upscaled to 1200x1200 for printing.
• Old scanned photos. 1990s scans at 300 DPI upscaled to modern print resolutions.
• Screenshots of high-DPI content captured at low DPI. UI mockups, screen grabs from older monitors.
• Video game screenshots or classic art. Neural upscalers do particularly well on stylized content because they can lean on training data for that style.
• Removing chunky pixelation from thumbnails you no longer have the source for.

Realistic non-wins:

• Reading text that's already too blurry to read. If a license plate is smeared beyond recognition, no upscaler will make it readable. The information isn't there.
• Faces in security camera footage. Same principle. Upscalers invent plausible facial features — they don't recover them.
• Compressed JPEG artifacts. Upscaling amplifies the artifacts. Denoise (Image Noise Reduction) before upscaling, not after.

Anything that requires forensic accuracy is beyond what upscalers can do honestly. Anything where "looks better" is enough — upscaling delivers.

Denoise before you upscale

The single most common upscaling mistake: taking a noisy or JPEG-artifacted source and upscaling. The upscaler faithfully enlarges the noise along with the content, and you end up with a bigger, sharper, more offensive version of the noise.

Correct order:

1. Image Noise Reduction first, to clean up compression artifacts and sensor noise.
2. Image Upscaler second, on the cleaned source.
3. Save the result at appropriate quality — see below.

Skipping step 1 wastes the upscaler's capacity on describing noise instead of content.

How much to upscale

Doubling (2x) is almost always safe — neural upscalers were trained heavily on 2x pairs and produce reliable results. Tripling (3x) works but shows more hallucination artifacts. Quadrupling (4x) is the practical limit before results start feeling "off." Beyond that, do multiple rounds (2x then 2x again) rather than one huge jump — the intermediate step gives the upscaler more to work with.

Save at the right format

Upscaled images are usually large. Two decisions:

• Keep as PNG if the content is graphics, text, or a screenshot. Preserves the sharp edges the upscaler worked hard to produce.
• Save as JPG at quality 85+ for photographic content. High quality avoids introducing new artifacts on top of the upscaled detail.

Don't save an upscaled image as low-quality JPG — you'll waste the whole point of upscaling.

Common gotchas

• Faces look plasticky. Some upscalers over-smooth skin. Try a different model if the tool offers one, or upscale less aggressively.
• Text looks slightly wrong. Upscalers can hallucinate letter shapes on very small text. If the exact letters matter (documents, receipts, screenshots), verify the output.
• Colors shift subtly. Some models normalize colors during upscaling. Compare side-by-side and adjust if needed.
• File sizes explode. A 500x500 → 2000x2000 PNG can go from 200 KB to 3 MB. Compress with Image Compressor if the target destination has size limits.

Doing it locally

Upscaling old family photos, ID scans, or medical images are all valid use cases where uploading to a cloud AI service feels wrong — because it is wrong. Image Upscaler runs the neural model in your browser via WebAssembly. Slower than a cloud GPU (10-60 seconds per image depending on size and your hardware), but the image never leaves your device.