How it works
The Real-ESRGAN model (SRVGGNetCompact, ×4) splits the image into tiles, runs each tile through the neural network on ONNX Runtime Web and stitches the result without seams. Pixels never leave the device. We do not call ordinary interpolation AI: if the model fails to load, the tool reports an error instead of substituting a result.
What to expect
The network works well on photos, screenshots and compressed images: it removes some artefacts and restores edges. It does not invent detail that was never there and does not replace professional restoration. On weak devices processing may take minutes.
Privacy
The file is read through the File API and processed in canvas. Network requests only fetch the model and runtime; the image is never uploaded. When the tab closes, the browser releases the memory.