AI Image Enhancer — Pro, on our server
A learned model is a set of expectations, and expectations only help where they fit. The clean-up network behind Cloud HD enhancement has seen an enormous number of photographs and almost nothing else, so it knows what a slightly soft eyelash, a compressed brick wall or a noisy patch of grass should look like when repaired. Hand it a screenshot of a spreadsheet and it has no comparable experience to draw on. This is not a flaw to be patched out; it is what makes the model good at the thing it is good at. Match the input to the prior and the output is convincing.
Text is the one failure mode that can pass unnoticed. Blur destroys the letter, the model rebuilds a letter, and the rebuilt one is chosen for plausibility rather than truth — so a licence plate, a serial number or a street sign can come back crisp and incorrect. Treat any reconstructed lettering as an illustration, never as a record. When what you actually want is the words rather than the picture, the on-device image-to-text reader extracts them in your browser and changes no pixels at all.
Artwork has a better answer than enhancement. A logo, an icon set or a line drawing is defined by hard edges and flat fills, and a photographic model tends to give those edges a slight organic wobble that reads as wrong on a brand mark. Rebuild the geometry instead: the image-to-SVG converter traces the shapes into a vector that is sharp at any size, on your device and free. For photographic images the opposite advice applies, and the sibling pages on improving photo quality and enhancing a scanned photo get more specific about each case.
The upload, in the exact terms it deserves. Cloud HD is Pro-only, opt-in, and the only part of MiniPx that leaves your device — every compressor, converter, cropper and resizer here works inside the browser. When you do run it, your image is uploaded to our server for that run and deleted when the run finishes, and in all cases within 24 hours; it is processed in the United States on Cloudflare R2 storage and Modal GPUs, encrypted in transit, and never used to train anything. The model that ran is named on the result — RestoreFormer++ today; the model can change, and you can always see which one ran.