AI image upscaler
Enlarge a small picture 2× or 4× with a model running on your device.
Files stay on your device. Nothing is uploaded.
Up to 10 MB and 250,000 input pixels (for example 500 × 500). PNG, JPEG, WebP or BMP.
The local engine and selected model download after you start. CPU processing can take a few minutes.
AI predicts detail and can introduce artifacts. Transparency is interpolated; check text, faces and edges.
Related tools
-
Image to text
Extract printed text from an image, then correct and copy it. -
Image compressor
Shrink photos to a quality level or to a target size, without leaving your browser. -
Image Resizer
Set the size, keep the proportions, download your images. -
AVIF to JPG
Turn AVIF pictures into JPG files, individually or in a ZIP.
Make a small picture larger using an image model rather than only stretching its existing pixels. Choose a twofold or fourfold enlargement and compare the result with the original before saving a PNG. The model runs locally and predicts detail; it does not recover a guaranteed true original. This first version is intended for small images, up to 250,000 pixels such as 500 by 500. Transparent areas are preserved separately from the color prediction, and your input file is never overwritten.
How to use
- Choose a PNG, JPEG, WebP or BMP up to 10 MB and 250,000 pixels in total.
- Select 2× or 4× and start processing; keep the tab open while the image is processed in small sections.
- Compare the labelled original and enlarged previews, inspect important details and download the PNG result.
What it handles
- Separate learned 2× and 4× models with real output dimensions, not a renamed resized file.
- Overlapping image sections bound memory use and provide neighboring context to the model.
- PNG export retains transparency; its alpha channel is interpolated separately from predicted RGB detail.
- Explicit input limit and cancellation keep expensive processing under your control; no image is uploaded.
FAQ
Are my files sent to an AI service?
Image enlargement predicts RGB pixels locally and handles transparency separately. Your file is processed in this browser, not uploaded to an inference service. Required models and runtime files are downloaded from this site after you start processing. The tool does not save your input or result in local storage. Keep the original file yourself and download any result you want to retain before closing the page.
Will upscaling recover the original details perfectly?
No. The model estimates plausible detail from the pixels available to it. Fine text, faces, logos and repeating patterns may change or develop artifacts, so inspect those areas before publishing. Increasing dimensions is not evidence of greater factual accuracy. Keep the original for comparison, especially when an image is used as a record or reference.
Why is there a limit of 250,000 input pixels?
Inference runs on your device's CPU in a worker, which avoids a paid service but can be slow. The limit bounds working memory and processing time; for example, 500 by 500 becomes 2,000 by 2,000 at fourfold enlargement. Larger input is rejected rather than silently reduced. Processing stops after five minutes; try a smaller image on slower devices.
Updated