AI Image Upscaler — Enlarge Photos up to 4×

Guide: enlarging small images with the AI upscaler

What happens inside the upscaler

The upscaler runs UpscalerJS on TensorFlow.js with an ESRGAN-slim super-resolution model. There is a separate model for each factor, 2x, 3x and 4x, with 4x selected by default, and the first time you use a factor its model, about 5 MB, is downloaded and cached by your browser. The network works through the image in small 64-pixel patches with a little overlap, so the joins do not show, and predicts sharper edges and texture than plain stretching would.

If your image is larger than 1200 pixels on its longest side, it is first scaled down to 1200 pixels, so a 1600 by 1200 photo is treated as 1200 by 900 and comes back at 4800 by 3600 at 4x. The result is a PNG named upscaled.png, and a note shows the size before and after. Your image is processed on your device and is not uploaded.

Picking 2x, 3x or 4x

Choose the factor from the size you actually need, not the largest one available. A 640 by 360 thumbnail becomes exactly 1280 by 720 at 2x. A 400-pixel logo that has to be 1200 pixels wide needs 3x. Higher factors take longer, use more memory and produce much larger files. When the size you need falls between two factors, upscale to the next one and then bring it down to the exact size with the Image Resizer; reducing a sharp image keeps it sharp.

Good candidates

• Small product photos from a supplier or a WhatsApp chat that look soft in a Flipkart, Amazon or Meesho listing.
• Old 640-pixel YouTube thumbnails or channel art.
• Profile pictures saved at a small size years ago.
• A person or object cropped out of a wider photo, such as one player from a cricket match photo.
• Simple graphics and logos that are needed at a larger size for a banner.

What it cannot do

The model predicts edges and textures; it does not know what a particular face, word or logo should look like. Small text can come out slightly warped, a face that is only a few pixels wide becomes a smooth larger face rather than a detailed one, and heavy JPEG blocks from a compressed image may be sharpened along with the real detail. Motion blur stays; for soft or shaky photos try the Blur Image Fixer.

Images that are already well above 1200 pixels gain less, because they are reduced before upscaling. If yours is already large and only needs to be a little bigger, the resizer may be enough.

File size after upscaling

A 4x result has 16 times as many pixels as the input, and PNG keeps all of them without loss, so the file can run to several megabytes. For websites, marketplace listings or forms, convert it with PNG to JPG or shrink it with Compress to Size.

Checklist for the best result

• Crop to the subject first with the Crop tool, so the model spends its effort where it matters and you stay under the 1200-pixel limit.
• Start from the least compressed copy you have.
• Use a stable connection for the first run of each factor.
• On a phone, close other tabs before using 4x.
• Compare before and after at full size, and keep your original.

Related guides

• How to Make a Small Image Bigger Without Losing Quality
• How to Make a Low-Quality Image HD Online
• How to Resize Images Without Losing Quality