About AI Image Upscaler
This image upscaler is free, needs no sign-up and runs in your browser, so your pictures are never uploaded. Choose a picture, pick x2 or x4, download the model once and press Upscale. You get a PNG that is two or four times wider and taller, with edges and texture redrawn by a super-resolution model rather than stretched.
The model is Swin2SR, a transformer built for image super-resolution, released under Apache-2.0. Each scale is a separate model of about 22 MB, fetched from Hugging Face the first time, with its size shown before you confirm. The model runs in a background worker, so the page stays responsive while it works.
Large images are handled in tiles. The picture is cut into 128 px squares that overlap by 16 px, each tile is upscaled on its own, and the overlaps are blended back together, so the whole image never has to fit in memory at once. That is also why there is a size limit: the result never goes past 4096 px on its long side, which means at most 2048 px in for x2 and 1024 px in for x4. A larger picture is shrunk to that limit first, and the tool tells you when it did.
Be realistic about what upscaling does. The model adds plausible detail, it does not recover detail that was never captured. It helps with small, soft or compressed pictures, old scans and low-resolution icons. It will not read a blurred licence plate or make a thumbnail look like a photograph.
How to use AI Image Upscaler
- Choose the scale: x2 or x4. Each has its own model of about 22 MB.
- Press Download model and confirm. The model comes from Hugging Face and your browser may keep it for next time.
- Add a picture by dropping it, tapping the box, pasting it or using the phone camera.
- Check the size note. If the picture is larger than the limit for that scale, it is shrunk first and a warning says so.
- Press Upscale. A progress bar shows the tile being worked on. Press Stop to cancel.
- Compare before and after with the slider on a phone or side by side on a large screen, then press PNG.
Features
- Swin2SR at x2 and x4, each about 22 MB.
- Tiled inference in 128 px tiles with a 16 px overlap that is blended, so big images do not run out of memory.
- Runs in a worker, with a progress bar and a Stop button.
- An honest size limit: the result stays within 4096 px on its long side, and the tool says when it shrank the input.
- Before and after compare, as a slider on phones and side by side at large widths.
- Download as a PNG at the upscaled size.
- Input by drop, file choice, paste or phone camera.
- Nothing is uploaded: only the model is downloaded.
How the tiles are planned
Tiles are 128 px squares. Neighbours share 16 px, so the start of each tile moves 112 px. The last tile in a row or column is moved back so it ends exactly at the edge of the picture. Where a tile overlaps the one before it, its opacity ramps up across the overlap, which blends the two and hides the seam.
step = tile - overlap = 128 - 16 = 112
starts = 0, 112, 224, ... and the last start = length - 128
output size = input size x scaleWorked examples
A 1024 by 768 px image at x4 gives 63 tiles
This is the largest input x4 accepts, because 1024 times 4 is 4096.
A 200 by 130 px image gives (0, 0) 128 by 128; (72, 0) 128 by 128; (0, 2) 128 by 128; (72, 2) 128 by 128
The second column of tiles starts at 72, not 112, so it ends exactly at the right edge.
A 96 by 64 px icon gives 1 tile of 96 by 64
An image smaller than one tile is processed whole.
Good to know
Upscaling is not the same as enlarging
Enlarging with bicubic or Lanczos resampling spreads the existing pixels over a bigger grid, which makes soft pictures bigger and still soft. A super-resolution model has learned what edges, hair and texture usually look like and draws them in, so the result looks sharper. Those details are a good guess, not a recording of the original scene.
For faces, text and logos, check the result closely. The model can invent strokes that were not there.
Speed
On a phone or a CPU-only browser a large picture can take minutes, because every tile is a full model run. A picture of 512 by 512 px at x4 is 25 tiles. If it is slow, try x2 or a smaller picture.
Privacy: how it works
- Pictures are processed in your browser and are never uploaded to Ulkaa or to anyone else.
- The only download is the Swin2SR model (about 22 MB per scale) from Hugging Face. That request does not contain your pictures.
Tips
- Start from the best source you have. Upscaling a heavily compressed JPG also enlarges its blocky artefacts.
- Use x2 first. It is about four times cheaper than x4 on the same picture.
- Crop the part you care about before you upscale, to spend the memory on it.
- Keep the original. The result is a new PNG and the tool never changes your file.
Frequently asked questions
Is it really free?
Yes. There is no price, no trial and no credit card. Ulkaa Tool does the work with your own device, so there is no server bill to pass on to you.
Do I need an account?
No. There is no sign-up and no login. Open the page and use it. An Ulkaa account is not needed for anything on this tool.
Does it run in my browser?
Yes. Your pictures are processed by code running in your browser tab, on your own device.
Is my data uploaded?
No. Your pictures are never sent to Ulkaa or to anyone else. You can check this yourself by watching the Network tab in your browser developer tools while you use the tool.
Is the image upscaled on a server?
No. The model runs on your device in a browser worker. Only the model file is downloaded.
Why is there a size limit?
A result of more than 4096 px on its long side can use more memory than many phones have. At x2 the input limit is 2048 px, and at x4 it is 1024 px. A larger picture is shrunk first, and a warning tells you.
What are tiles?
The picture is cut into 128 px squares that overlap by 16 px. Each is upscaled on its own and the overlaps are blended, so memory use stays low whatever the image size.
Will it make a blurry photo sharp?
It makes a small or soft picture larger and adds believable detail. It cannot recover detail that was never captured, such as a blurred number plate.
Which model does it use and what is its licence?
Swin2SR, in its classical x2 and x4 versions, licensed Apache-2.0. Each is about 22 MB.
Can it do x8?
No. You can run the result through again, but quality drops and the 4096 px limit applies to the output.