AI Image Upscaler
Create a true 2× super-resolution image locally with a Swin2SR neural network.
Instead of stretching pixels like a normal resize, this tool runs the Swin2SR super-resolution neural network through Transformers.js, downloaded and executed entirely inside your browser via WebGPU or WASM. It reconstructs plausible detail at 2x resolution rather than just interpolating existing pixels, which is why small product photos or old low-res images come out noticeably sharper than a canvas-based upscale. Inputs are capped at 10 MB, 1024 px per side, and 400,000 total pixels; for bigger images or when speed matters more than AI detail, use the standard Image Upscaler instead.
Loading tool
How it works
- 01
Drop or select a JPG, PNG, or WebP image up to 10 MB and 1024 px.
- 02
Wait while the Swin2SR model downloads and caches on first run.
- 03
Download the 2× super-resolution result.
Details
- Category
- AI
- Runs
- In your browser
- Files uploaded
- No
- Cost per run
- Free
- Sign-in
- Not required
- Works offline
- Yes, once loaded
Last updated 2026-08-03
Common questions
Is this really AI upscaling?
Yes. It runs the Swin2SR super-resolution neural network through Transformers.js, rather than stretching pixels with canvas interpolation.
Why is the first run slow?
Your browser downloads and caches quantized model files once. Inference then runs locally with WebGPU when available, or WASM otherwise.
What are the limits?
Inputs can be up to 10 MB, 1024 px on either side, and 400,000 total pixels. The model always outputs at 2× resolution.
How does this differ from the plain Image Upscaler?
The plain Image Upscaler uses fast canvas interpolation with no size cap. This tool runs an actual neural network for sharper reconstructed detail, at the cost of stricter input limits and a slower first run.