What an image enhancer changes, and what it leaves alone
Every run of this image enhancer makes the picture bigger: the output is 2x or 4x the processed width and height, and the new pixels are filled with detail the model learned in training rather than stretched from the old ones. You get crisper edges on shapes and lettering, less haze from earlier resizing, and fewer of the square blocks and speckles that heavy JPEG compression leaves around text and flat colour. Scanner and sensor noise is reduced too, though rarely removed completely.
Just as important is what the image enhancer leaves alone. Layout, wording, colours and every object in the frame come through as they were, so a watermark, a date stamp or an unwanted logo is sharpened along with everything else rather than removed. Nothing is colourised; greyscale stays greyscale. Detail that was never captured cannot be recovered: a tagline only a few pixels tall turns into a cleaner smudge, or into shapes that look like letters but are not. Motion blur is softened at best, never fully undone.
Choosing a mode: one engine for most files, Anime for flat art
The mode menu offers four options, but there are really two engines. Basic, Photo and Screenshot all run the same model, the x4plus version of Real-ESRGAN on fal.ai, with face enhancement off; the names are presets, not separate models. If Photo and Screenshot return near-identical results for the same file, that is expected. For anything that is not flat colour, such as photographic texture, soft gradients or scanned paper, stay with one of these three.
Anime mode is the genuine alternative. It runs a different Real-ESRGAN model, RealESRGAN_x4plus_anime_6B, which was trained on anime art and suits flat colours and clean line art: logos, icons, cartoons, simple charts and drawn meme templates. For a mixed image, such as a flyer that combines a photo with a flat logo, run it at 2x in both engines and check each result on the before/after slider, and decide at 2x before committing to 4x.
Logos and icons: when the source file beats enhancing
A logo is the clearest case where an image enhancer should be the fallback rather than the first move. Logos are usually drawn as vectors, so an SVG, EPS or PDF version, or the original design file, often exists somewhere, and it can be exported at any size with exact colours and clean edges. Look on the brand's press or media page and in shared drives, or ask whoever designed it. A fresh export beats any upscaled bitmap, and print shops often ask for vector artwork anyway.
When no source file exists, choose between redrawing the mark and enhancing the bitmap. A simple logo with a few flat colours can be redrawn or auto-traced in a vector editor, and enlarging the small file 4x in Anime mode first can give the tracing tool smoother edges to follow. Icons work the same way: a 32×32 favicon enlarged 4x is only 128×128, and most of what you see at that size is the model's guess, so recreate icons meant for app stores or large tiles instead of stretching tiny ones.
Memes, thumbnails and images from old sites or documents
Memes and reaction images are forgiving, since the goal is a clean look at a bigger size rather than accuracy. The weak spot is baked-in caption text: after years of screenshots and re-uploads the letters are blocky, and enhancing firms up those damaged edges rather than redrawing the font. If you know the words, enhance the picture and set the caption again on top.
Before running a thumbnail or an image lifted from an old website through an image enhancer, look for a larger copy. Older sites often link each thumbnail to a bigger file, and web archives sometimes keep the original upload. For pictures inside documents, extract instead of taking a screenshot: a .docx file is a zip archive whose word/media folder holds the embedded images at their stored size, and most PDF tools can export embedded images directly. A 300-pixel-wide thumbnail at 4x becomes 1,200 pixels wide, which suits an article body but not a full-width banner.
Saving the enhanced file: PNG, JPG or WebP
Pick the format for where the image is going, not the one it arrived in. PNG is lossless, which makes it right for logos, icons, diagrams, anything with text or flat colour, and for a master copy you plan to edit again; the cost is large files for photo-like content. JPG suits photographic images on the web at a high quality setting, but it is a poor home for crisp graphics, because saving an enhanced logo as a low-quality JPG puts blocks and halos back around every edge.
WebP can be lossy or lossless and usually gives smaller files than either, which suits web pages, though some print shops, older programs and upload forms still reject it. JPG cannot store transparency: a logo or icon that needs a transparent background belongs in PNG or WebP, checked against both light and dark backgrounds, and the site's background remover can cut a graphic out of a solid backdrop as a separate step.
Using an image enhancer for the web versus for print
For the web, an image enhancer should aim at display width multiplied by pixel density. An image shown 600 pixels wide in a page layout needs 1,200 real pixels to look sharp on a typical 2x high-density phone or laptop screen, which is exactly what a 300-pixel source gives at 4x. Going far beyond the display size only adds download weight. Once the graphic is on the page, the site's alt-text generator can draft a description for it.
Print asks more of an image enhancer. About 300 pixels per inch is the usual target for flyers and anything read at arm's length, and around 150 is often acceptable for posters seen from across a room. A letter-size flyer (8.5 × 11 inches) at 300 ppi needs 2,550 × 3,300 pixels; a scan of 850 × 1,100 pixels reaches 3,400 × 4,400 at 4x, which covers it. Keep the ceiling in mind: in Basic, Photo and Screenshot modes, anything above about 2.25 megapixels is scaled down in the browser before a 4x run (6 megapixels before 2x), so their 4x output peaks around 36 megapixels.