Why small faces in a wide frame go soft
How sharp a face looks depends mostly on how many pixels it covers. In a close headshot the face fills much of the frame, so eyelids, lashes and the rim of each iris get plenty of pixels. In a wide shot or a group photo the sensor spreads its pixels across the whole scene, each face gets a small patch, and fine features merge into soft shapes.
Sharing shrinks faces further, since many apps downscale and recompress photos. A face upscaler can only work with the pixels it receives, and this site adds one more squeeze: before a 4x run, any image larger than about 2.25 megapixels is scaled down to that size in your browser (6 megapixels for 2x). A face that was already small gets smaller still, so how you frame the upload matters more than any setting.
Crop close before you run the face upscaler
Cropping is the most useful thing you can do before a face upscaler runs. Picture a 4032×3024 phone portrait in which the face covers about 600×750 pixels. Uploaded whole for 4x, the photo is reduced to about 1732×1299, the face drops to roughly 258×322 pixels, and 4x only brings it back to around 1030×1290. Crop the same photo to 1000×1250 around the head first: at 1.25 MP it is under the limit, nothing is reduced, and 4x returns a 4000×5000 image built from every original pixel of the face.
Leave some margin. Hair, ears, the jawline and a little background give the model context, and a slightly loose crop is easier to reframe later as a round avatar. Any crop tool will do; save the crop as a copy and keep the untouched original.
Group photos: one face or one row at a time
Group shots are the hardest job for a face upscaler, because every face shares the same pixel budget of about 2.25 MP at 4x. In a wide family or team photo, each face may be only a few dozen pixels across after the reduction, where eyes and mouths are hints rather than details. Upscaling the whole picture still cleans it up overall, but it will not make every face crisp.
For the faces that matter, crop them out separately, one person or one row of three or four per file, and you get a larger, cleaner version of each for an album or a staff page. Each crop is a separate job: everyone gets 2 free credits a day, 2x costs 1 credit, 4x costs 2, and credit packs start at $9.99.
2x or 4x for profile pictures, headshots and prints
Profile pictures are mostly seen small: a round avatar in a chat, a thumbnail beside a comment, a square on a team page. For those, 2x is usually plenty. What matters is a clean face at the size people actually see, and the site you upload to will usually shrink a very large file anyway.
In the face upscaler, pick 4x when the image needs room: a headshot for a printed program or poster, a portrait you plan to crop again, or a large photo on a website. With a very soft source, try both scales on the slider; 2x often looks more natural because invented texture is packed into fewer pixels.
Selfies in low light: noise and phone smoothing
Indoor and evening selfies are usually noisy, and phones fight that noise with strong smoothing, which can leave skin looking like watercolor and turn hair into one soft mass. The face upscaler reduces noise and JPEG blocks to a degree and firms up edges, so a grainy selfie usually comes back cleaner. Texture that the phone already smoothed away, however, does not return as real detail.
Upload the original from your camera roll rather than a screenshot or a copy saved from a chat, since each extra round of compression removes more of the face. Skip beauty filters and heavy sharpening beforehand; they remove or exaggerate texture that then gets enlarged. A very dark portrait full of colored speckles will come back smoother and slightly flat rather than with crisp skin.
Eyes, hair and skin: what a face upscaler can sharpen
Eyes draw the most attention, and results from any face upscaler vary most there. When the eyelids, the iris edge and the small highlight in each eye are visible in the source, they usually come back crisper. When each eye is only a few pixels wide, the model draws a plausible shape instead, which can look slightly off, so check the eyes at 100% zoom before posting.
Hair often shows the clearest gain: its outline sharpens and strands become more distinct, though flyaways that were already blurred may turn into generic texture. Skin tends to come out smooth, since reducing noise and compression blocks can also soften very fine texture, and pores or fine lines only reappear where the source recorded them. Moles, freckles, scars and wrinkles stay; this is not a retouching tool.
Why identity stays intact: face generation is off
Some face tools add a second model that detects faces and repaints them from what faces generally look like. The result can be sharp but slightly different: another nose, other eyes, a younger look. In this face upscaler that step, GFPGAN face enhancement, is switched off. Photo, Basic and Screenshot all send your image to the same Real-ESRGAN model, which treats the face like the rest of the picture.
The benefit is that the person stays recognizably themselves: expression, asymmetry, marks and age are left alone. The honest cost is that it will not rebuild a face that is only a blur. A smudge of a few pixels becomes a larger, cleaner smudge, and heavily blurred features stay approximate; a tool that produces a sharp face from almost nothing is generating that face, not recovering it.
Official photos, other people and honest use
Some uses require an unaltered photograph, such as passport, visa, driving license and ID card applications. A portrait that has been through a face upscaler is an AI-processed image, so check the rules of the authority before you submit one, and if in doubt use the original or have a new photo taken. Likewise, because the model can invent plausible texture, a sharpened face is not proof of who someone is.
If a photo shows other people, ask before you enhance and share it. A larger, sharper face is easier to recognize. For team pages and school photos, check with the people involved, or with parents where children appear.