1. Home
  2. Images
  3. AI Image Upscaler

AI Image Upscaler

Enlarge low-resolution photos 2× or 4× with an AI model running locally.

Runs in your browser · No upload

How to use

  1. Drop the photo to enlarge onto the box or click to choose it. You can also paste a screenshot with Ctrl+V.
  2. Choose the scale (2× or 4×) and the method (AI or Fast). The AI input size, result size, number of tiles and estimated time appear right away.
  3. Click Start enhancing. With the AI method, the model (7.1 MB) is downloaded and saved on this device the first time only. You see progress and time left, and Cancel stops it immediately.
  4. Hover over the result to see that spot at 2×, the original and the result side by side. Choose a save format (PNG, JPG or WebP) and Download.

Methods and options

Item Description
AI 2× Runs the lightweight Swin2SR model once. AI input up to 1,000px on the long side; result up to 2,000px
AI 4× Runs the AI a second time on the 2× result. Input up to 500px on the long side; result up to 2,000px
Fast No AI: the browser’s high-quality interpolation, then sharpening. It finishes quickly but cannot create new detail
Faces · photos / Art · screenshots Both choices use the same AI model; only the final sharpening (unsharp mask) differs
Reduce JPEG artifacts Applies a light blur first so blocky compression patterns are not enlarged. Fine texture softens a little too

Why it takes time, and how it is estimated

The AI model needs about 4.5 seconds per 128px tile on an N100 mini PC. So the photo is split into 128px tiles that overlap their neighbors by 16px; after processing, the overlaps are mixed with a linear weighted average so there are no visible seams. The estimate is the number of tiles × 4.5 s, and it is recalculated from the real speed once the first tile is done.

Input Scale Tiles Result Estimated time (N100)
256×256 2× 9 512×512 45 s
640×480 2× 30 1280×960 2 min 15 s
1000×1000 2× 81 2000×2000 6 min 5 s
320×240 4× 36 (9 + 27) 1280×960 2 min 45 s
500×500 4× 106 (25 + 81) 2000×2000 8 min

Larger photos can be processed by turning on the option that reduces them to fit the limit. For example a 3000×2000 photo is reduced to 1000×667 for 2× and to 500×333 for 4× (the result gets smaller by the same amount). If you need a larger result, use the Fast method, which reduces the input only enough to keep the result within 8,000px per side and 16 million pixels.

Honest limits

AI enlargement “invents plausible” detail that is not in the photo, using patterns it has learned. Texture and outlines may improve, but text, fine facial detail and repeating patterns can end up different from the real thing. This is a small, light model, so its quality does not match professional upscalers, and a very blurry or heavily compressed original may not improve much. For anything important, such as printing or records, check the result by eye. The model is Swin2SR lightweight x2 (quantized), licensed under Apache-2.0.

Privacy and storage

The photo is processed only in your browser (in a web worker) and is never sent to a server. The model file is downloaded from this site and kept in browser storage (Cache Storage); its SHA-256 is checked when downloaded to reject damaged files. You can delete it any time with “Delete saved models”.

To only make an image smaller or larger by size, use Resize Image; to remove a background, use Background Remover; to remove unwanted objects, use Remove Objects from Photo.

Your photos are processed in your browser and never sent to a server.

Can't find the tool you need?

Tell us what you want to do. We build the most requested tools first.