AI Image Upscaler
Enlarge photos, artwork and anime stills by 2x or 4x with tiled inference so large images stay fast and sharp instead of blurry.
Preview
How to Upscale an Image
Add your image
Drag and drop one or more PNG, JPG, WebP, GIF or SVG images into the upload area, or click to browse your device. Each file can be up to 500 MB.
Tip: you can add several images at once — the whole queue is upscaled with the same settings.
Pick a model and scale
Choose one of the 7 on-device AI models, then set the output to 2x or 4x. Nothing is uploaded.
Upscale and download
Press Upscale image and watch the preview update. Download files one by one, or grab everything at once as a ZIP archive.
Frequently Asked Questions
How much can I enlarge an image?
Each model has a native factor, and you can also force 2x or 4x. 4x on a 2000px image produces an 8000px result.
Which model should I choose?
PurePhoto for realistic photos, Real-ESRGAN and LSDIR for mixed content, Anime for illustrations, and Lite when you want the smallest, fastest model.
Does upscaling add detail or just stretch pixels?
It runs a neural network that reconstructs edges and texture, so results look sharper than simple resizing. Very low-resolution sources can still look soft.
Are my images uploaded to a server?
No. The model runs locally in your browser through ONNX Runtime Web, so your images never leave your device.
Which file formats are supported?
PNG, JPG, WebP, GIF and SVG are accepted, and transparency is preserved in PNG and WebP results.
Is the alpha channel kept when upscaling?
Yes. The original transparency is preserved, so logos and cut-outs stay transparent after upscaling.
Can I upscale a whole batch at once?
Yes. Drop any number of images into the queue, process them all together, then download each result or everything as a ZIP.
How large can the source images be?
Each file can be up to 500 MB. Extremely large photos are refused with a clear message before the AI starts.
Why is the first run slower than the next ones?
The AI model is downloaded once and cached in your browser, so the first run is slower and every later run starts immediately.
What is the difference between the model native factor and the 2x/4x scale?
The scale forces the output size, while each model has a native factor it was trained for. 4x models look sharpest at 4x, 2x models at 2x.