Enlarge a small image without it turning to mush
A logo saved at 300 pixels wide, or an old photo that falls apart when you print it. A neural upscaler invents plausible detail rather than just stretching pixels, and the same class of model that desktop software charges a subscription for runs here.
- Model
- Swin2SR classical super resolution, x2
- Scale
- 2x in one pass, 4x by running it twice
- Runs on
- Your GPU via WebGPU, falling back to CPU
- First use
- A one time model download of 52 MB, then cached
- Best on
- Inputs under about 1200 pixels on the long edge
How to use it
- Drop in the small image. Under 1200 pixels on the long edge is the sweet spot.
- Pick 2x or 4x. Four means two passes, so it takes about three times as long.
- The model runs on your machine. The first run downloads the weights once.
- Compare it against the original and download the PNG.
Questions
Is this the same as Real-ESRGAN or the paid desktop tools?
Same family of idea, a smaller model. It is genuinely better than the plain stretching your image editor does, especially on edges and text. It will not match a 200 MB desktop model on a badly damaged photograph.
Why is there a size limit on the input?
Super resolution work grows with the pixel count, and a large input in a browser tab will run out of memory. Above roughly 1200 pixels there is also little point, since you already have the detail.
Can it un-blur a photo?
Not really. It adds plausible detail at higher resolution, which is a different problem from recovering detail that motion blur destroyed. Expect an improvement, not a rescue.
Does it invent things that were not there?
Yes, that is how these models work, and it is worth knowing before you use one on anything evidential. Faces and text are where invented detail shows up most.
Next door
heic to jpg, compress pdf to a size, remove background, passport photo.