Upscaling increases an image's resolution. Ordinary resizing (bilinear or bicubic interpolation) only spreads the existing pixels out, which makes edges soft and blurry. This tool instead uses a super-resolution neural network that was trained on real photos to predict sharper edges and finer texture at the higher resolution.
Your everyday toolkit
Image ToolsImage Upscaler
Upscale images 2× or 4× using browser-based super-resolution. Improve image detail and sharpness while keeping your photos private on your device.
A little less effort
Image Upscaler
Upscale images 2× or 4× with super-resolution
Your next creation starts here.
Drop your image here
or drag and drop it here
Using Image Upscaler well
Enlarge a small photo using a locally executed super-resolution model and compare the result with the source.
Example to try
Use 2× on a 500 × 400 image to produce 1000 × 800 pixels. Examine faces and small lettering at full size before using the export.
What to check before relying on the result
Predicted texture is not recovered evidence. The model can invent details or smooth text. Model/runtime downloads need a connection; image size and device memory limit processing.
How to use Image Upscaler
View largerStarting exampleLow-resolution source
80 × 100 pxYOU’RE EXPLORING
Upload your image
A few steps.
A whole new result.
Frequently Asked Questions
The useful details, before you get started.
Real-ESRGAN General x4v3 — the compact general-purpose model (SRVGGNetCompact, about 1.2 million parameters) released by the Real-ESRGAN authors (Wang et al., 2021) under the BSD-3-Clause license. It is a trained 4× super-resolution network, run in your browser with ONNX Runtime Web on your GPU (WebGPU) when available, otherwise on your CPU (WebAssembly).
No. Processing happens entirely in your browser. The only thing downloaded is the model file itself (about 4.9 MB, fetched once from Hugging Face and then cached by your browser) and the ONNX Runtime engine from the jsDelivr CDN. Your image never leaves your device.
2× doubles the width and height (4× the pixel count), e.g. 1000×800 becomes 2000×1600. The model works natively at 4×; for 2× the 4× result is downsampled by exactly half with area averaging — the same approach Real-ESRGAN's own tools use. So 2× takes about as long as 4× but gives a smaller file.
4× quadruples the width and height (16× the pixel count), e.g. 500×400 becomes 2000×1600. It is the model's native scale and is best for small images, thumbnails, and low-resolution photos.
Partly. It sharpens edges, reduces JPEG blockiness and blur, and can produce plausible texture, but it cannot recover information that was never captured. Fine detail it adds is the model's best guess: faces and small text can look slightly painted or smoothed, and very noisy or heavily compressed images may come out over-smoothed. Always check the result before using it where accuracy matters.
JPG, JPEG, PNG, and WebP up to 20 MB. Results can be downloaded as PNG (lossless), JPG, or WebP. Transparent PNGs are supported: the colors are upscaled by the model and the transparency channel is resized with standard high-quality resampling, because the model only handles color.
The model processes the image in 128×128 pixel tiles (with a small overlap so seams don't show), and larger images need more tiles. Speed depends heavily on your device: with WebGPU a typical 1 megapixel photo takes seconds to tens of seconds; on CPU only it can take a minute or more. The output is limited to 36 megapixels so the browser doesn't run out of memory.
Choose any question to see the details.

