AI Image Upscaler — Upscale Images 2× or 4× in Your Browser
Enlarge a small or low-resolution image 2× or 4× with an AI super-resolution model that runs on your device. Compare before and after, download PNG, JPG or WebP.
Found a problem, or something that could be better? We read every message and we fix things quickly.
This free AI image upscaler enlarges a small or low-resolution image to two or four times its size with a super-resolution neural network, so edges stay sharp and textures look like texture instead of the soft blur an ordinary resize gives. Drop in a photo, a product shot, an old web image or a screenshot, choose 2× or 4×, and compare the result against the original with a slider before you download it as PNG, JPG or WebP. Several images can be upscaled in one batch (up to 10, one after another) and downloaded together as a ZIP.
The model is Real-ESRGAN (general x4v3), an open-source model trained to undo the blur, noise and JPEG artefacts of real-world low-quality images, and it runs entirely on your device: the image is never uploaded. Be clear about what it does, though. An upscaler cannot recover detail the original never captured; it predicts plausible detail from patterns it learned. That is ideal for making a small image look clean at a larger size, and wrong for anything where the exact detail matters, such as reading a blurry licence plate or identifying a face.
How to use
- Drop an image on the box or click to choose one (JPG, PNG, WebP and other common formats). For a batch, select or drop up to 10 images at once.
- Choose 2× or 4×. The output size is shown; images whose 4× version would exceed 8192 px per side or 16.7 megapixels must be resized smaller first.
- Press Upscale. The first run downloads the AI engine and model, then the image is processed in tiles with a progress bar. Cancel at any time.
- Drag the comparison slider to check the result, and switch to View at 100% to inspect details at full size.
- Pick PNG (lossless, keeps transparency), JPG (smaller, white background) or WebP (smallest) and download. In a batch, download each image from its row or all of them as one ZIP.
Frequently asked questions
How does AI image upscaling work?
A super-resolution model is a neural network trained on millions of pairs of sharp images and degraded, smaller copies of them. It learns how edges, textures and fine structures usually look at high resolution, and when it sees a small image it predicts the missing pixels from those patterns. A normal resize only interpolates between existing pixels, which is why it looks soft; the AI result looks sharper because it adds detail, but that detail is a learned guess, not information recovered from the original.
Is my image uploaded?
No. The model runs in your browser using WebGPU or WebAssembly, and the image stays on your device. The only downloads are the AI engine and the model file on first use, served from this site and then cached by your browser. That makes it safe for unreleased product photos, client work and personal pictures.
Should I choose 2× or 4×?
Choose the smallest scale that gives you the size you need. 4× turns a 500 px image into 2000 px and is the model’s native scale. 2× is the same 4× result scaled back down with high-quality resampling, which usually looks a little cleaner and gives a file a quarter of the size. Both take the same processing time.
Why is it slow on my computer?
Upscaling runs billions of calculations per image. With WebGPU (current Chrome and Edge on most desktops and laptops) the graphics chip does the work and a typical image takes seconds. Without WebGPU it runs on one CPU core: in our tests a 300 × 225 px image took about 10 seconds on a recent laptop, a 500 × 500 px image takes roughly half a minute to a minute, and older computers and phones can be several times slower. The progress bar shows an estimate after the first tiles; you can keep using other tabs while it works.
What kinds of images upscale well?
Photos, product images, illustrations, logos and small web images that are reasonably in focus upscale well, and the model also cleans up JPEG blockiness and mild noise. Heavily blurred or motion-blurred photos stay blurry, small text can come out distorted, and faces may look smoothed or slightly altered. Always check the result at 100% before using it.
Can I upscale multiple images at once?
Yes, up to 10 images per batch. They are upscaled one after another on your device, with a progress bar for the current image and one for the whole batch; each finished image can be compared and downloaded on its own, or all of them saved as one ZIP in the format you pick. Because AI upscaling is demanding, a batch can take from under a minute with WebGPU to much longer on a CPU. Cancel stops the remaining images and keeps the finished ones.
Is there a size limit?
The upscaled result can be up to 8192 px on its longest side and 16.7 megapixels in total (about 4096 × 4096), which keeps memory use within what browsers allow. For 4× that means inputs up to roughly 1024 × 1024 px; upscalers are meant for small images, and a photo that is already large rarely benefits. The file itself can be up to 25 MB.
AI upscaling vs ordinary resizing
Every image editor can make an image bigger. Bicubic or Lanczos resizing calculates each new pixel as a weighted average of its neighbours, so a 4× enlargement contains exactly the same information spread over sixteen times the pixels: edges turn into soft gradients and JPEG artefacts get magnified. It is predictable and faithful, and for small enlargements (up to about 1.5×) it is all you need; the image resizer does it instantly.
An AI upscaler replaces the averaging with a prediction. Real-ESRGAN was trained on real-world degradations (blur, noise, compression, resizing), so it produces crisp edges, removes much of the blockiness and invents plausible fine texture. The comparison slider shows the difference directly: the left side is the original stretched by the browser, the right is the model’s output.
The trade-off is fidelity. Because the added detail is generated, the result can contain things that were never in the photo: smoothed skin, regular patterns where the original had noise, letters that look sharp but are slightly wrong. For a website hero image, a product thumbnail or a print of a small photo, that is a good trade. For evidence, documents or anything where accuracy matters, keep the original.
Getting the best result
• Start from the best source you have. Upscaling the original file beats upscaling a screenshot of it or a copy that has been re-saved several times.
• Crop first. Upscaling only the part you need is faster and keeps the output within the size limit; the image cropper does that without uploading anything.
• Use 2× when it is enough. A smaller enlargement asks the model to invent less, and the file is a quarter of the size of 4×.
• Check at 100%. Fit-to-width hides artefacts; switch to View at 100% and look at faces, text and fine patterns before you publish.
• Compress afterwards. A 4× PNG can be tens of megabytes; for the web, download WebP or run the result through the image compressor.
About the model and your privacy
The tool uses realesr-general-x4v3 from the Real-ESRGAN project by Xintao Wang and colleagues, released under the BSD 3-Clause licence. It is a compact model (about 5 MB, 1.2 million parameters) designed to handle general real-world photos rather than only clean test images, which is why it suits a browser: it is small enough to download quickly and fast enough to run on a laptop. It is executed with ONNX Runtime Web, Microsoft’s open-source inference engine.
Large images are split into overlapping tiles, each tile is upscaled separately, and the centres of the tiles are stitched together, so memory use stays bounded and no seams are visible. Transparent images keep their transparency: the model works on the colour channels, and the alpha channel is enlarged with high-quality resampling and put back. All of this happens in your browser tab; close it and nothing remains anywhere.
Related free tools
- Image Resizer — Shrink images by percentage, exact pixel dimensions, or a target file size like 500 KB — all in your browser.
- Image Compressor — Shrink images by quality or to a target size, optionally cap the dimensions, and download smaller JPG or WebP files. Batch, in your browser.
- Background Remover — Remove the background from a photo with an on-device AI model. The image never leaves your computer. Download a transparent PNG or fill a colour.
- Image Cropper — Crop images to any size or aspect ratio in your browser — square, 16:9, 4:3, or freeform.
- Image Converter — Convert between AVIF, HEIC, TIFF, PNG, JPG, WebP, GIF, BMP and SVG. Most conversions never leave your browser.