AI image upscaler
Upscale a photo 2× or 4× with an AI super-resolution model — entirely in your browser, no uploads.
Runs 100% in your browserOriginal
Choose an image.
Upscaled
Result appears here.
Both previews are shown scaled to fit — download the PNG to see the upscaled image at full size.
How to upscale an image with AI
- Choose an image. Select a JPEG, PNG or WebP photo of up to 16 megapixels.
- Pick 2× or 4×. Choose how much bigger you want it — 2× for already-sharp photos, 4× for small or soft ones.
- Upscale and download. Run the model in your browser and save the enlarged image as a PNG.
What AI super-resolution actually does
Enlarging a picture used to mean interpolation: to invent the pixels between the ones you have, average the neighbours. That is why a stretched photo looks soft — nothing new is ever added, the existing detail is just spread over more pixels. Super-resolution takes a different route. A small neural network is trained on millions of pairs of the same image at low and high resolution until it learns what a downscaled edge, a strand of hair or a patch of fabric usually looked like before it shrank. Given a new picture it predicts that missing structure rather than averaging: edges come back crisp, texture comes back as texture. This tool runs one such network — a compact Real-ESRGAN "general" model — tile by tile across your image, which is how a few megabytes of weights can upscale an image far larger than the network was designed to see in one pass.
When to upscale, and when to just resize
Reach for AI when the source is the limiting factor: an old scan or family photo, a thumbnail that is the only copy left, a heavy crop from a phone shot, a product shot that a marketplace wants at a larger size, or artwork you want to print bigger than its native resolution allows. Those are the cases where you need to genuinely increase image resolution, not just change the dimensions. If your original is already sharp and large enough and you only need different numbers — a 1200 px hero, a 512 px avatar — a plain resize is the better tool: it is instant, it is exactly faithful to the original, and no model can improve on detail that is already there. The image resizer handles that case, and it is worth trying first.
Honest limits
An upscaler predicts plausible detail; it does not recover the detail that was really there. That distinction matters. Motion blur, heavy JPEG artifacts and genuinely low-resolution sensors destroy information, and no amount of computation brings it back — the model fills the gap with what images like yours usually contain. On landscapes, fabric, foliage and general texture the guess is convincing and the result looks like a better photo. On faces, small text, logos and fine line work the guess is more visible: skin can look waxy or painted, letters can turn into confident nonsense. So treat the output as a new rendering rather than a restored original, always compare it against the source at full size, and never use it where the exact pixels matter — documents, evidence, anything that will be read rather than looked at. For the same reason, upscale once from your best original instead of running the result through again, and if the goal is simply to enhance photo quality for the web, the image compressor will shrink the (rather large) PNG this tool produces without undoing the work.
Why it runs in your browser
Nearly every free service that offers to upscale image files does it server-side: you upload the photo, their hardware runs the model, you download the result — and your picture, which may be a face, a document or unreleased work, has spent time on someone else's machine under someone else's retention policy. Here the model itself is downloaded to you instead, and it runs on your CPU (or your GPU via WebGPU, where the browser supports it) as WebAssembly. The photo never travels. The trade-offs are honest ones: a one-time download before the first run, a size ceiling set by your device's memory, and processing that takes seconds rather than being instant. In exchange the tool keeps working with the network off, and nothing you feed it is ever logged. The background remover works the same way if you need to cut the upscaled subject out afterwards.
Frequently asked questions
- No. The super-resolution model runs inside your browser — the image is read from your device, processed there, and written back there. Nothing is uploaded, nothing is stored, and no account is needed. Almost every other free AI upscaler sends your photo to its servers to do the work; this one does not.
- A plain resize stretches the pixels you already have and interpolates between them, so a blurry photo simply becomes a bigger blurry photo. An AI upscaler runs a neural network that has learned, from millions of image pairs, what fine detail typically looks like at higher resolution — so it reconstructs edges, texture and grain instead of smearing them. If all you need is different dimensions from an already-sharp image, the plain image resizer is faster and more faithful.
- Use 4× when the source is genuinely small or soft — an old scan, a web thumbnail, a heavy crop — and you want the biggest possible reconstruction. Use 2× when the source is already reasonably sharp and you just need more resolution for print or a retina screen: it runs the same 4× model and then halves the result, which keeps the learned detail while cutting the file size and hiding any over-sharpening. When in doubt, 2× looks natural more often.
- The AI model (about 5 MB) and the WebAssembly runtime that executes it have to be fetched once before anything can run locally. Both are cached hard by your browser afterwards, so every later image starts instantly — and, once cached, the tool keeps working offline.
- Three things happen only once: the model downloads, the runtime is compiled, and the model is warmed up on the first tile. After that you are paying for computation alone. Repeat runs on the same page are noticeably quicker, and a browser with WebGPU support will use your graphics card, which is several times faster than the CPU path.
- Yes — 16 megapixels (roughly 4000 × 4000). The limit exists because everything happens in your device's memory: a 16 MP photo at 4× already means a 256-megapixel intermediate image, which is close to the largest single canvas a browser will allocate. If your file is bigger, crop or resize it first, or choose 2× rather than 4×.
- You can load JPEG, PNG or WebP. The result is always a PNG, so the upscaled detail is stored losslessly with no second round of compression artifacts. PNGs of this size are large by nature — run the output through the image compressor if you need it for the web rather than for print.
- No, and be sceptical of any tool that claims otherwise. The model produces plausible detail, not the detail that was actually in the original scene — information genuinely lost to blur, heavy JPEG compression or a low-resolution sensor cannot be recovered. It excels at texture, edges and general sharpness; faces, small text and logos can come back looking slightly painted or invented. Always compare against the original before using the result.