marcus.li ~/upscale
built with claude nothing uploaded runs entirely in your browser

the lab

Free AI Image Upscaler

Upload a low-resolution image and get a genuinely sharper, larger version back, using a real AI super-resolution model, not a stretched resize. It runs entirely in your browser after a one-time model download; your image is never uploaded anywhere.

why this isn't just a resize

Stretching a small image with a resize (even a good one like Lanczos) just spreads the same pixels over more space, so it always looks soft. This uses a small convolutional AI model trained on real photos to reconstruct plausible edges and texture as it scales the image up, which is why the result looks noticeably sharper, not just bigger.

It's not magic: it can't invent detail that was never captured, so a very blurry or heavily compressed source will still look soft at a larger size, just less blocky than a plain resize would look.

1 · your image
Click to choose an image, or drag and drop it here.
JPG, PNG, or WebP. Best results on photos rather than screenshots of text.
2 · scale
First use downloads the AI model (a few megabytes, one time, then cached by your browser). Processing happens on your device, so a phone or an older laptop will be slower than a modern desktop; very large images are automatically capped before processing to avoid crashing the tab.
Loading AI model…
how this works
  1. The AI model (an RDN-based super-resolution network trained on the DIV2K photo dataset) and TensorFlow.js are loaded into your browser, from this site, not a third-party service.
  2. Your chosen image is decoded locally and, if it's very large, automatically scaled down first so the model can process it without exhausting your browser's memory.
  3. The model processes the image in small overlapping tiles, reconstructing detail tile by tile, since running the whole image through at once would need more memory than most browsers allow.
  4. The tiles are stitched back into one final image, which is displayed next to the original and offered as a download. None of this touches a server.