AI Image Upscaler – Enlarge Photos 2× or 4×
Upscale and enhance photo quality with AI that runs on your device. Free, no watermark, nothing is uploaded.
Processed on your device · no uploadDrop images here or choose files
JPG, PNG, WebP · up to 50 MB each · runs on your device, nothing is uploaded · or paste with Ctrl+V
The AI option downloads a small model once (about 5 MB) and keeps it for offline use. Without WebGPU the instant Sharp resize is the default.
What to do next
How to use AI Image Upscaler
Add your images
Drop one image or many. Sharp resize is instant; AI adds more detail (Photo for photos and screenshots, Anime / art for drawings).
Choose 2× or 4×
Press Upscale. The first run downloads the AI model (about 5 MB, once); after that it works offline.
Compare and save
Drag the slider to compare before and after, then download PNG, JPG or WebP, or all results as one ZIP.
Frequently asked questions
Is my image uploaded?
No. The AI model downloads once and runs on your own device. Your pictures never leave it, and the tool keeps working offline after the first run.
Which AI model does this use?
Photo mode uses Real-ESRGAN realesr-general-x4v3 (BSD-3-Clause, by Xintao Wang et al., github.com/xinntao/Real-ESRGAN, ONNX export from vlad-grig/realesr-general-x4v3-onnx, 4.9 MB). Anime / art mode uses the official Real-ESRGAN realesr-animevideov3 weights (BSD-3-Clause, release v0.2.5.0 of the same repository, SHA-256 b8a83768…18b75d) in the ONNX export skillsafe-ai/realesr-animevideov3, 2.5 MB. Both run with onnxruntime-web (MIT). The files are fetched from Hugging Face once and cached on your device.
What is “Sharp resize” and when is it used?
A fast non-AI method: Lanczos enlarging plus gentle edge-aware sharpening. It takes about a second and needs no download. Browsers without WebGPU (Safari, Firefox, many phones) would need minutes per megapixel for the AI, so Sharp resize is the default there and the AI is an option with a time estimate. It is cleaner than normal enlarging, but it cannot invent new detail the way the AI does.
Does it use the GPU?
The AI uses WebGPU when your browser has it (Chrome and Edge on a computer today), which is roughly ten times faster than the CPU path. Without it the AI runs on the CPU, single-threaded.
How big can the picture be?
The limit is the size of the result: up to 64 megapixels on a computer, 32 on a computer with little memory, 16 on a phone (8 on an older one). A 2000 × 1500 photo at 4× is 48 MP. When an image is too big for 4×, the tool tells you to use 2×. It works best on a computer; WebGPU is much faster than the CPU path.
Will it make a blurry photo sharp?
It adds believable detail and cleans up soft edges and JPEG blocks, but it cannot recover what was never captured. Faces and small text can look slightly different from the real thing, so check the result before using it for anything important.
Why is a smooth image not “better” by every measure?
Real-ESRGAN is built to look sharp rather than to match pixel for pixel. On detailed images (text, screenshots, textures) it scores higher than plain bicubic enlarging; on very smooth gradients the two are close and it may add faint texture.
Does it keep transparency?
Yes for PNG and WebP: the alpha channel is enlarged smoothly. JPG has no transparency, so transparent areas become white.