Local and private
Your source image and upscaled output stay on your device. Only static model assets are downloaded.
Three super-resolution models · self-hosted by WKO AI
Enlarge images 2x or 4x without uploading them. Pick a model for speed, clean detail, JPEG artifacts, or noisy real-world photos; inference runs locally and the weights come from our own domain.
Try it freeOpen Upscale and choose a small image first. Your source stays in the browser. Larger images need more working memory.
Start with Lightweight 2x for speed. Choose a 4x model for more output pixels. Check download sizes below and device compatibility in the studio.
Approve the first model download, run the upscale, and compare the result before saving. CPU fallback may take longer.
Actual studio results
We reduced the 512×512 generated example to 128×128, then processed that input through both upscalers in WKO AI. Each preview uses the same display area; the source is enlarged for comparison.
Reduced input
Download PNG
Swin2SR Lightweight 2x
Download PNG
Real-ESRGAN General x4v3
Download PNGUpscaling changes texture and sharpness; it cannot guarantee recovery of original detail. Compare the full-size PNGs before choosing a result. These examples have no additional sharpening or retouching.
Choose the right restoration
Every shipped weight carries a permissive license (Apache-2.0 or BSD-3-Clause), is SHA-256 verified during mirroring, sits below Cloudflare's 25 MiB asset limit, and is cached by your browser after the first use.
| Model | Scale | Format | Size | Best for |
|---|---|---|---|---|
| Swin2SR Lightweight 2xdefault 2x | 2x | fp32 | 8.08 MB | Speed on any device |
| Real-ESRGAN General x4v3default 4x | 4x | fp32 | 4.87 MB | Photos, screenshots, JPEGs |
| Swin2SR Classical 4x | 4x | q8 | 21.6 MB | Clean-source 4x fidelity |
Your source image and upscaled output stay on your device. Only static model assets are downloaded.
No credits, account, watermark, or server-side processing bill. Your device performs the inference.
Each model page includes exact bytes, SHA-256, license attribution, and the same ONNX file used by the studio.