Photo Enhancer

No. Everything here is measurement and correction: the histogram, the colour statistics and the noise level are measured, and the image is adjusted accordingly. That is a real and important distinction. An AI enhancer generates detail that was never captured — it can look spectacular, and for some purposes it is exactly what you want. But it is inventing. This redistributes information that is genuinely in your file, which is why the result is safe to use for a document, a product photo, or anything that might be taken as a record of what was actually there.

Not really, and you should be sceptical of anything that claims otherwise. Sharpening increases contrast at edges, which makes a slightly soft image look crisper, and there is a sharpening control here that does that well. But motion blur and missed focus destroy information, and no amount of correction brings it back. What this fixes properly is exposure, colour, contrast, haze and noise — problems where the detail is present but poorly presented.

Use Shadows rather than Exposure. Exposure lifts everything uniformly, which brightens the dark areas but also pushes the already-correct midtones and highlights toward white — that is what makes a brightened photo look flat and grey. Shadows applies a curve weighted to the dark regions only, using a large blurred mask so it works on areas rather than individual pixels. The Too dark preset combines a modest exposure lift with a strong shadow recovery and some noise reduction, because shadow detail is always the noisiest part of an image.

It removes a colour cast. Indoor tungsten light is strongly orange, fluorescent tends green, and open shade is blue — your eye adapts and stops noticing, but the sensor records it faithfully, which is why photos taken indoors so often look yellow. The tool estimates the colour of the light falling on the scene and scales each channel to neutralise it, using the Shades of Grey method — a middle ground between assuming the average of the scene is neutral and assuming the brightest pixel is white, both of which fail on their own.

Because it cannot tell the difference, and guessing would ruin the photo. Estimating the colour of light is what is called an ill-posed problem: a grey wall photographed under orange light and an orange wall photographed under white light produce genuinely identical pixels. There is no algorithm that separates them, only guesses. What breaks the tie in ordinary photographs is variety — a real room contains several surface colours, so a bias shared by all of them is probably the light rather than the objects. When the whole frame sits in one narrow band of hue, as a sunset or an underwater shot does, that evidence is missing, so the automatic correction stands down rather than draining the colour you wanted. The slider is still there, and the estimate is still shown: you can see it is a sunset, and the tool cannot.

Removing the atmospheric veil that flattens distant subjects. It uses the dark channel prior, which rests on a genuinely clever observation: in a clear outdoor photograph almost every small patch of the image contains something very dark in at least one colour channel — a shadow, a gap, a dark surface. Haze adds light uniformly, which lifts that per-patch minimum. Measuring how much it has been lifted estimates how much haze sits in front of each region, and that can be subtracted. It is the same technique used in serious photo software.

Only partly, and it will tell you when it cannot. If a region is pure white in all three channels, the information is gone — there is nothing to recover, and no tool can do better. The Highlights control works on areas that are bright but not yet clipped, pulling them back down and restoring the tonal separation. If your image has a significant proportion of pure white pixels, the analysis says so explicitly rather than implying it can be undone.

Saturation scales every colour equally, which is why pushing it makes already-vivid areas go neon while the dull parts still look dull, and why it wrecks skin tones first. Vibrance weights the adjustment toward colours that are currently muted and leaves saturated ones alone, with an additional guard on the warm, moderately saturated range where skin sits. For most photographs vibrance is the one you want; saturation is there for when you know you want the blunt version.

It can if you push it, which is why it is set conservatively and why the slider is there. The filter is edge-preserving: neighbouring pixels only contribute to the average if their brightness is close to the centre pixel's, so smoothing happens inside flat regions and stops at boundaries. That keeps edges intact where a plain blur would destroy them. But every noise reduction removes some genuine fine texture along with the noise, and past a certain point skin and foliage start to look waxy. Compare with the split before you commit.

No. The analysis and every correction run in your browser on your own machine. Nothing is transmitted, no copy is kept, and once this page has loaded it works with the network off. For family photos, documents or anything you would rather not hand to a stranger's server, that is the whole point.

Yes — drop in as many as you like and they are all analysed on load. You then choose how the batch is treated, and the choice genuinely matters. "Auto per photo" measures each image separately and corrects it on its own terms, which gives the best result for any individual picture. "Same for all" copies one correction from the photo you are looking at across the whole set. Neither is right in every case: correcting individually is wrong for a sequence shot under one light, because a correction that varies shot to shot makes an obviously inconsistent set; copying one is wrong when the photos genuinely differ. A batch downloads as a ZIP; a single photo downloads on its own.

There is no fixed limit, but the work happens on your machine, so a large batch of large photos takes real time and memory. They are processed one at a time at full resolution and each working canvas is released before the next begins, which is what stops a long batch exhausting memory part way through. If one file cannot be decoded it is marked unreadable and skipped rather than failing the whole run.

It should not, and it will not in any way you can see. The preview runs on a copy fitted to about 900 pixels so that dragging a slider is instant rather than taking several seconds, and downloading re-runs the identical pipeline on your full-resolution original. Every stage is resolution-independent except the two spatial radii — the shadow mask and the haze map — which are derived from the image dimensions so they scale proportionally.

A great deal, and it is fixed for good reasons. White balance comes first, because every later judgement is made on the corrected colours. Dehaze comes before levels, while the veil is still a simple additive term. Shadows and highlights follow, once the global range is settled. Noise reduction comes after the tonal moves that would otherwise have amplified noise invisibly, and before sharpening, which always runs last — sharpening anything you subsequently blur or stretch is wasted work.

Yes, and there is a preset for it. It pushes the levels hard so paper goes properly white and text goes properly black, fully neutralises the colour cast that scanners and phone cameras give paper, drops the saturation so leftover colour fringing disappears, and sharpens more aggressively than you would want on a photograph. The result is markedly more readable and much smaller as a JPEG.

It reads anything your browser can display — JPG, PNG, WebP, GIF, AVIF and BMP — and writes JPEG, PNG or WebP. JPEG is the default here because enhanced photographs are usually destined for sharing; choose PNG if the image has transparency or you intend to edit it further.