AI Image Upscaler
Upscale images 2× with AI super-resolution — powered by Swin2SR, runs entirely in your browser.
Drop an image or click to browse
PNG, JPG, WebP — best results under 512×512 px
Upscaled result will appear here
AI Image Upscaler — 2× Super-Resolution in Your Browser
This tool uses Swin2SR, a Swin Transformer V2-based image super-resolution model, to double the resolution of images directly in your browser. Swin2SR achieved state-of-the-art results on benchmark datasets when it was published and produces sharp, natural-looking output without the ringing artefacts common in older upscaling methods. The quantised model (~150 MB) downloads once and is cached locally. Works best on images up to 512×512 pixels — larger inputs work but take longer. The output is always a lossless PNG at exactly 2× the input dimensions. Nothing is uploaded to a server.
Built and maintained by Meet Shah · Last updated
What this tool is used for
- Enlarging a small image where a plain resize would look soft.
- Recovering usable detail from a low-resolution asset.
- Upscaling a screenshot for a print or a large display.
- Getting a larger version of an image whose source has been lost.
- Comparing an AI upscale against a plain resize.
Frequently Asked Questions
- Does upscaling recover lost detail?
- No — it invents plausible detail. A model trained on many image pairs predicts what a higher-resolution version would look like, which is a very different claim from recovering what was there. The result can look sharper and be wrong about specifics.
- How is it different from bicubic resizing?
- Conventional interpolation averages neighbouring pixels, which is why an enlarged image looks soft. A learned model instead generates edges and texture that were never in the source, so it looks sharp — at the cost of hallucinating detail that is not evidence.
- Where does that matter?
- Anywhere the pixels are evidence. Upscaling a licence plate, a face from CCTV, or a document scan produces a confident image that is a guess, and it has been misused as if it were enhancement. For illustration and print it is fine; for identification it is not.
- Why is the first run slow?
- Because the model weights download and compile before anything can be processed — tens of megabytes, once, then cached. That cost is what buys running locally, so the image is never uploaded anywhere.
- What kind of image upscales best?
- Clean sources with clear structure — line art, logos and well-exposed photographs. Heavily compressed JPEGs upscale badly because the model faithfully amplifies the block artefacts, treating compression noise as detail worth reconstructing.
Common errors and gotchas
- Expecting real detail to be recovered, when the model is inventing plausible detail rather than restoring it.
- Upscaling faces or text, where invented detail is most obviously wrong.
- Assuming the first run works offline, when the model downloads once.
- Upscaling a lossy image, where the model amplifies compression artefacts along with the subject.
- Using an upscale where the correct fix is to obtain the original at full size.