Sharpening the whole frame, not just a face
This page runs a deblur pass over the entire image, so hedges, fabric weave, wood grain, product labels, printed text, and background objects come back at the same time as the people in the shot. That is the practical difference between it and the free face restoration tool, which runs a face-specific model: it finds faces, rebuilds eyes, teeth, hairlines, and skin, and leaves everything else in the picture untouched. If the only soft part of your photo is a face and the rest of the scene already looks fine, pick face restoration, because a dedicated face model reconstructs facial features more aggressively than a general deblur pass can. If the whole shot is soft, because focus missed, the lens was smeared, or the file came back mushy from a chat thread, pick this page. Running the same photo through both and keeping whichever result reads better is a perfectly reasonable way to decide.
It is also worth separating unblurring from upscaling, since the two get mixed up constantly. The free AI image upscaler makes an image bigger, adding pixels so a small file can fill a screen or survive a print. Unblurring makes a soft image sharp at the size it already is. Enlarging a blurry photo on its own just gives you a larger blurry photo, so the two compose in one direction: unblur first, then upscale the sharpened result.
What comes back
The model is the Qwen image editing model with a deblur adapter trained specifically on the blurred-to-sharp pairing, rather than a general purpose editor being asked politely to sharpen things. That matters, because a generic edit instruction tends to add contrast and edge halos and call it detail. A trained deblur adapter reconstructs the texture that the blur smeared out, which is a different operation from raising local contrast.
A test on a garden portrait makes the scope clear. The input had been through a gaussian blur strong enough to flatten the whole frame. The output brought back leaf structure in the hedge behind the subject, a visible weave in the shirt, separated slats on the wooden bench, and facial detail, all in one pass, with nothing masked or selected by hand. That whole-frame behaviour is the reason this tool exists next to the face-specific one: the parts of a photo that sell it as sharp are usually the small textures everywhere else, not only the face.
The composition is held fixed. Subject placement, colors, crop, and framing come out as they went in, so the result stays usable as a version of your photo rather than a reinterpretation of it.
The honest limit
Moderate softness is where this works well. Slight focus miss, camera shake at a shutter speed that was a stop too slow, a cheap lens, heavy compression, an old phone camera, a screenshot of a screenshot: all of those hold enough residual structure for the model to rebuild from, and the improvement is obvious.
Extreme motion blur is a different case, and so is a photo that has degraded past the point where you can tell what you are looking at. Those cannot be reconstructed. The detail is simply not present in the file to recover, and no model can retrieve information that was never recorded. What a model can do in that situation is invent something plausible, which is worse than leaving the photo alone, because an invented face or an invented license plate looks confident and is wrong. Treat any result as a sharpened version of your photo, not as evidence, and be especially careful with identity and text in badly degraded inputs.
Where it fits, and what to expect
For a full cleanup pass, order helps. Unblur first, then enlarge: the free AI image upscaler handles straight enlargement, and the free Clarity AI upscaler adds invented detail while it enlarges, which suits photos you want to look richer rather than merely bigger. Scanned prints with scratches, creases, and fading are a job for the free old photo restoration tool before anything else, since damage repair and deblurring are separate problems. And if a face is still the weak point after a deblur pass, finish on the free face restoration tool.
The tool accepts JPG, PNG, and WebP up to 4096px per side, and re-renders at roughly one megapixel matching your input's aspect ratio, so a sharpened file comes back around that size regardless of what you uploaded. If you need it larger, run the upscaler on the output. Everything runs on shared free GPUs, so a pass takes around fifteen to thirty seconds depending on queue depth. It is free, there is no sign-up, and there is no watermark on the result. Your upload lives on the GPU only while the photo renders, then the transient files expire. Nothing is saved to a gallery and nothing is used for training.