Tiny, fast and auditable
The pinned 109 KB AMD weight is Apache-2.0 licensed and verified by exact SHA-256 during every build.
Upscale an image with real local AI
Recover cleaner edges with a real super-resolution network, then drag across the exact same pixels to judge it against ordinary browser resizing. The image never leaves this device.
↗SESR-S109 KBLOCAL AI
Drop a low-resolution image hereCtrl/⌘ V or paste an imageJPEG, PNG or WebP · up to 50 MB
SESR-S INT82× or two-pass 4×
First use loads the 109 KB model and local ONNX runtime. Used assets are cached for later offline reuse.
SESR-S INT8 · Apache-2.0
Judge the recovered detail
Browser resizeLocal AI
Model inference, comparison and export stay in this browser. No request contains the image.
Continue locally
Apache-2.0 · INT8 · ONNX
A measurable neural result, not an “AI” label on interpolation
The tool runs a pinned INT8 SESR convolutional network on RGB tiles and shows ordinary high-quality browser resizing underneath the same split view. The receipt exposes inference time, tile count and mean neural pixel delta so you can decide whether the model actually helped.
The baseline stays visible
Fit, 100% and 200% inspection modes compare the neural output with ordinary resizing at identical dimensions.
2× and honest 4×
The 4× option is explicitly two successive 2× neural passes, with a 16 MP output guard instead of hidden cloud quotas.
Local AI upscaling questions
Is this really AI, or just resizing?
It runs the SESR-S INT8 convolutional super-resolution model through ONNX Runtime. The before side is browser resizing, and the receipt reports how much the neural pixels differ.
Does the image get uploaded?
No. Only the model and runtime are downloaded. Image decoding, neural inference, comparison and export all happen locally.
Why can 4× look too strong?
4× applies the 2× model twice. Some sources benefit; others develop ringing or invented texture. Lower AI strength or use 2× when the comparison looks less natural.