Enhance text in an image without rewriting it
Text is where upscalers get caught. A slightly invented leaf is invisible, but a slightly invented letter turns a word into nonsense. That is why this page is a separate preset of the free AI image upscaler: the general tool leads with a diffusion model that rebuilds texture and can redraw small print into convincing lookalike glyphs. This page offers the same three models, but it defaults to GAN, which is faithful. GAN and Real-ESRGAN enhance text in an image by sharpening the strokes that are already there, keeping the font, the stroke weight, and the contrast against the background, while multiplying the resolution 4x.
If what you really need is the words themselves as editable text, not a sharper picture of them, use the free image to text tool instead, or upscale here first and then extract, since cleaner strokes help text recognition on small print.
Make blurry text readable: how to use it
Upload a screenshot, a scan, a photo of a sign or a whiteboard, or an infographic, in JPG, PNG, or WebP up to 2048 px per side. Press Sharpen Text and the result appears on a comparison slider at 4x the size. Zoom into the smallest line of text: that is the honest test of any upscaler.
GAN is the default and gives the crispest letter edges, especially on small sources. Real-ESRGAN is worth a second try when the text sits in a compressed JPEG, because it was trained to remove the fuzzy halo that compression leaves around sharp edges, which is often what makes small text hard to read. Seed VR 2 Small is the third option: it can help on a very soft or heavily compressed source, but it can redraw letters and always returns about 2 megapixels, so treat it as a last resort for lettering. Face restoration is switched off on this page, since it has nothing to do with letters.
Lossless sources give the best results. A PNG screenshot beats the same screenshot re-saved as a JPEG, and a scan made at a higher resolution beats a phone photo taken at an angle. When you must work from a photo, crop to the text area first so the size limit is spent on the letters.
Photos of documents benefit from a little preparation. Straighten the page so lines of text run horizontally, crop away the desk and your fingers, and use the brightest, most even light you can when you take the picture: shadows across a page lower the contrast of the letters, and an upscaler sharpens contrast that exists rather than creating it. For receipts and thermal paper, which fade fast, take the photo or scan as soon as possible. For a multi-page document, upscale the pages one at a time, then assemble them into a PDF with your usual tool. A sharp 4x page also makes a much better printout than the original small image, since small text no longer prints as grey smudges.
Sharpen text in screenshots, scans, and signs
Screenshots for documentation, slides, and bug reports are the easiest case: they are already sharp at their own size, and a 4x upscale keeps menus, code, and small UI labels crisp when shown large on a projector or a high-density screen. Scanned contracts, receipts, and archive pages gain cleaner, darker strokes that are easier to read and to print. Photos of signs, storefronts, and menus come out with lettering that holds up when zoomed. Infographics and charts that were reshared until they went soft get their labels back.
Before you publish or file anything, proofread it. GAN and Real-ESRGAN do not swap characters, but when the source letters were ambiguous, a smudged 8 that could be a 6, the upscale will sharpen the ambiguity rather than resolve it.
What it cannot recover
An upscaler can only sharpen information that is still in the pixels. If the letters are distinguishable in the source, even barely, they will come back clearer. If compression or blur has already merged them into blobs, no faithful model can tell which letters they were, and a model that claims to would be guessing. Text a few pixels tall in a wide photo, such as a distant license plate or a page seen across a room, is usually beyond recovery.
For best results, start from the original file, avoid repeated screenshots, and keep the downloaded result as is rather than re-saving it as a low-quality JPEG. Images are processed transiently and are not stored.
If a screenshot only needs to be bigger for a slide, the 4x result can be scaled down to the exact size you need in any editor and will still look sharper than the original enlarged directly.