AI Image Upscaler
Rebuild detail at 2× or 4× with an AI model that runs on your device — no upload, account or watermark.
- Runs in your browser — files never leave your device
Drop an image here, click to choose, or paste
JPG, PNG or WebP · one image at a time · phone photos welcome
Your image and the AI model stay in this browser. Nothing is uploaded.
Best for small or low-resolution images. Photos larger than 2.5 MP (2×) or 1 MP (4×) are reduced first, and the AI works in tiles of up to 256 px.
How to use
- Drop, choose or paste one JPG, PNG or WebP image.
- Pick 2× for a quicker result or 4× for the largest AI reconstruction.
- Choose PNG, JPEG or WebP, then select “Upscale with AI”.
- Watch the model and tile progress, compare Before and After at 100%, then download the result.
AI super-resolution, not ordinary resizing
A normal resizer interpolates existing pixels. This tool sends the image through Real-ESRGAN General x4v3 in tiles of up to 256×256 pixels, a super-resolution neural network, then stitches only the non-overlapping tile cores together. Each tile has 16 pixels of surrounding context to reduce visible seams. 2× output is produced from the AI-generated 4× image with a high-quality downscale, so both choices use the model rather than plain enlargement.
Model, privacy and limits
The self-hosted model is realesr-general-x4v3, an ONNX export of Real-ESRGAN General x4v3. Its 4,868,454-byte (4.87 MB) model file is distributed under BSD-3-Clause; it and ONNX Runtime load only after you start a job. The image never leaves this tab. Inputs up to 50,000,000 pixels, 16,384 px per side and 50 MB are accepted. To bound memory and processing time, the AI works on a proportional copy of at most 2.5 MP at 2× or 1 MP at 4×; the notice above the button shows the exact output dimensions before you start.
Set realistic expectations
Upscaling predicts plausible texture; it cannot retrieve text, facial features or object detail that was never captured. At 4×, faces, lettering and repeating patterns can look painted or invented, so inspect the 100% comparison before using a result for print or evidence. The AI works on at most 2.5 MP at 2× and 1 MP at 4× (larger photos are reduced first), so a typical photo needs about 60 model tiles: a WebGPU-capable desktop normally feels much quicker than a phone using the WebAssembly fallback. Start with 2× on a phone or when the original is already fairly clear.
Frequently asked questions
Are images uploaded to a server for AI upscaling?
No. The image, 4.87 MB Real-ESRGAN model and ONNX Runtime run in the current browser tab. Your source image is not sent to Betterbit or a third-party AI service.
What is the difference between 2× and 4× image upscaling?
2× doubles width and height, so it creates 4 times as many pixels. 4× quadruples width and height, so it creates 16 times as many pixels and takes more memory and processing time.
How long does AI upscaling take?
Processing time depends on the device, browser and image size. This tool shows model-download progress, each inference tile and an estimate of the time left; photos are reduced to at most 2.5 MP (2×) or 1 MP (4×) first, so a typical photo needs about 60 tiles. Desktop WebGPU is generally several times faster than phone WebAssembly. You can switch tabs while it works.
Does image upscaling recover unreadable text or blurred faces?
No. AI super-resolution estimates likely detail from nearby pixels, so it cannot guarantee factual recovery of unreadable text, heavily blurred faces or missing objects. Inspect the result at 100% before relying on it.
Does the tool keep transparent PNG images transparent?
Yes. PNG and WebP results preserve separately resampled transparency. JPEG does not support alpha, so transparent areas are flattened to white.
Related tools
Updated 2026-10-28