Image upscaler

How to upscale a picture

1. Add your picture. Drop it anywhere on the page, paste it, or click the box. JPEG, PNG, WebP and iPhone HEIC photos work.
2. Choose 2×, 3× or 4× larger. A line under the buttons says the exact new size, for example 400 × 250 px becomes 800 × 500 px, and how many tiles the work is split into.
3. Check the picture type. Automatic picks a model for photos or for anime and drawings. Change it if the guess is wrong.
4. Press Make it larger. A progress bar shows the tiles being finished, with a time estimate once it has measured its own speed. You can cancel at any time.
5. Compare and download. Drag the slider to see the original, simply enlarged, next to the result. Switch View to Actual size to inspect the detail, then press Download.

What it does, and what it does not

Stretching a picture to twice its size, as the image resizer does, cannot add detail: it only makes the existing pixels bigger and smoother, so the result looks soft. An upscaling model has learned from many pairs of sharp and blurry pictures what edges, hair, fur, leaves and line art usually look like when they are sharp, and it draws that detail in as it enlarges. The result looks sharper and cleaner.

It is important to understand that the added detail is a plausible guess, not a recovery of what the camera or the artist saw. In practice:

• Textures, edges and line art improve most. Fur, hair, grass, brick, fabric and the outlines in anime art come out crisp.
• Small details can change. Letters on a sign, a pattern on a shirt or the exact shape of a small face can come out differently from the original. Do not use it where the exact content matters, such as evidence, documents, license plates or scientific images.
• Compression damage can be sharpened too. A heavily compressed JPEG with blocky patches gets cleaner, but strong blocks can turn into odd textures.
• It cannot rescue a tiny picture. A 32-pixel icon enlarged four times is still guesswork on top of guesswork. Larger starting pictures give better results.

The two models

Both are small models from the open-source Real-ESRGAN project by Xintao Wang and colleagues, released under the BSD 3-Clause license, which allows this use. Each is a 4× model, saved in a format that runs in browsers, from Hugging Face, and pinned to one exact version. Your browser keeps the model after the first download, so the next picture starts at once.

Automatic uses the anime model when a single exact color covers at least 40% of the picture (checked on a small copy), as in a drawing, a logo or a screenshot with large flat areas, and the photo model otherwise. Artwork with a detailed background can fail that test, so pick the type yourself if the note names the wrong one. Photographs almost never have one exact color over that much of the frame, because noise and light spread the colors out. The note under the picture type says which model is in use.

2×, 3× and 4×

Both models enlarge four times. To make 2× or 3×, each finished piece is shrunk from 4× to the size you chose, which gives a sharper, cleaner picture than asking a model for less. A 2× result therefore takes as long as a 4× one, so choose the size you need.

Added detail

The Added detail slider mixes the model's result with a plain smooth enlargement of the same picture. At 100% you get the model's full result. Lower it if the picture looks too sharp, too processed or too busy, which can happen with faces or with soft artwork. At 0% the model has no effect and you get an ordinary enlargement.

How it works, and how we checked it

A picture is cut into tiles of 192 pixels, and each tile goes to the model with some of its neighbors around it, 24 pixels for the photo model and 16 for the anime model, so that the model sees what is next to the edge. The extra border is cut off again afterward, and each tile is drawn straight into the finished picture. Several tiles are worked on at the same time, up to four, using your computer's spare processor cores.

We compared the tiled result with the same model run over a whole picture at once, using an independent program. Inside the picture the two agree to within 0.3 levels out of 255 on average, and across a tile join the difference is no bigger than beside it, so you should not see seams. The padding is what makes that true: with only 8 pixels of context, the differences were up to about 15 levels in our tests.

Pictures with see-through areas

A PNG or WebP with transparency keeps it. The see-through map is enlarged smoothly on its own, and the colors under the see-through parts are first spread outward from the nearest solid pixels so that the model does not draw a dark or white fringe around the cutout. A picture with see-through areas is saved as a PNG unless you pick another type, because JPEG cannot be see-through.

Size limits and time

• The result can be up to 64 megapixels and 16,384 pixels on a side. Choices that would be larger are greyed out, and the page tells you when a picture is already too big for any enlargement. In that case, shrink it or crop the part you need first.
• Time grows with the number of pixels. A 400 × 250 picture took about four seconds on our test computer with four workers. A 3000 × 2000 picture is 176 tiles, which takes a few minutes. Phones and older computers are slower, and use fewer workers.
• Memory. The finished picture is held in memory until you save it, so a very large result can fail on a phone.

Saving

• Same type as the original keeps JPEG as JPEG, PNG as PNG and WebP as WebP (PNG for see-through pictures). iPhone HEIC photos become JPEG. You can also choose PNG, JPEG or WebP, and set the quality (50 to 100, starting at 92) for JPEG and WebP.
• The file is named after your picture with the scale added, for example holiday-2x.png.
• Metadata is not kept. The saved file is a new picture, so camera details and GPS location are left behind. See our EXIF guide. The phone's rotation tag is applied first.
• HEIC photos work. In browsers other than Safari, the page downloads a decoder of about 3 MB the first time you add one.

When it helps

• Printing or displaying a small picture larger, such as an old photo or a web-size image.
• Wallpapers and game or video assets made from low-resolution art.
• Anime and illustration that only exists as a small file.
• Training sets. If a collection has some low-resolution images, enlarging them can bring them to a common size, but remember that the extra detail is generated. Our guides on image size and count for LoRA training and the dataset workflow cover size choices, and the bulk image resizer brings a folder to one size afterward.

Privacy

The model runs inside this tab, and your original file on disk is never changed or deleted. The page downloads the model from Hugging Face the first time you use it. For what the site collects, see our privacy policy.