Image Enhancer

Most photographs are wrong in a small number of predictable ways: the exposure is off by a fraction of a stop, the white balance drifted indoors, the shadows are crushed or the highlights are blown, and the tonal range never reaches either end. Fixing that by hand means knowing which of those applies before you touch anything — which is why a slider box is a tool for people who already know what is wrong with their picture. This one measures first and then acts, and it shows you the measurement so you can disagree with it.

What you get

• A correction derived from your photo, not a fixed recipe applied to every photo. The same scene setting produces different numbers for a dark picture and a bright one — that is what makes it automatic rather than a preset.
• Reasons, not just results. Every decision is shown with the measurement behind it: how many stops the exposure was out, what fraction of pixels were clipped, what the noise floor measured. No competitor shows you the inside of its enhance button.
• Every decision is adjustable. The numbers land on real sliders. Move one and it is yours; the others stay automatic, including across a batch.
• Nothing is uploaded. The photograph is read, measured and corrected on your device. No queue, no credit meter, no watermark.

1. 1Add your photo: Drop a JPEG, PNG, WebP, AVIF, BMP or GIF onto the page. It is read on your device and never uploaded.
2. 2Read what was measured: The panel lists what the tool found — how far the exposure is off, how much colour cast is present, how much of the tonal range is unused, the noise floor and whether the picture is in focus.
3. 3Pick a scene if the default is not right: Auto suits most photographs. Portrait protects skin, Low light lifts and denoises harder, Scanned drives paper to full black and white, Faded restores a print that has lost its blacks.
4. 4Adjust anything you disagree with: Every decision is a number on a slider. Move one and it becomes yours; the rest stay automatic. Reset to auto puts them all back.
5. 5Enhance and download: Press Enhance for a full-resolution result. With several photos, each is corrected from its own measurements and the whole batch downloads as one ZIP.

Six things are measured before anything is changed, each with a published estimator rather than an invented one — because every number is shown to you, and a made-up number presented as a measurement is worse than no number at all.

What is measured

• Tonal range. Per-channel and luminance histograms, and the levels at which the darkest and brightest half-percent of the picture sit. This is what says whether contrast is missing.
• Exposure, in linear light. Brightness is averaged in linear light, not in the stored values, because a ratio between two exposures is only meaningful there. A correction derived from the stored midpoint over-corrects dark pictures.
• Colour cast. Two independent estimates — one that assumes the average of the scene should be neutral, one that assumes the brightest near-neutral surface should be — fused by how much they agree. Where they disagree the scene probably just contains a lot of one colour, and the correction is reduced rather than guessed.
• Noise floor. A high-pass estimator that cancels smooth gradients exactly, sampled across nine full-resolution patches and read at the quietest of them. Noise is a floor: every patch shows its own noise plus its own texture, so the least textured patch is the closest reading.
• Focus. The variance of the Laplacian at full resolution, read at the sharpest patch. A photograph with one crisp subject against a blurred background is a sharp photograph.
• Haze. The dark-channel prior, normalised by the scene's own atmospheric light, with a confidence that falls to nothing when the scene has no bright region for the model to apply to.

What is corrected

• Levels, three different ways. Auto contrast stretches luminance only and cannot move a hue. Auto tone stretches each channel independently for maximum range. Auto colour does that and pulls each channel's midpoint to neutral, which is what actually removes a cast. Most tools pick one of these silently; this one lets you choose.
• Exposure and recovery. Exposure is corrected toward a target measured in linear light and damped, so a deliberately moody photograph is nudged rather than normalised. Highlight and shadow recovery are driven by how much is actually at the rail, not by where the tail sits — a bright sky with nothing clipped needs no recovery, and pulling it down anyway is how automatic tools produce flat grey results.
• Clarity. Midtone-weighted local contrast at a large radius, applied as a luminance shift rather than a gain so it does not double as a saturation control. Weighted away from both ends, because contrast added to a near-black region has nowhere to go but the clip.
• Denoise, coupled to sharpening. The measured noise floor sets the denoising strength and simultaneously caps the sharpening, because sharpening is a high-frequency gain and noise is high-frequency. Deciding them independently is what produces the over-processed look: smeared texture and crunchy grain at once.
• Dehaze. A transmission estimate from the dark-channel prior, inverted to recover the scene behind the haze. Off unless you ask for it — haze correction on a photograph with no haze is only a contrast boost with a longer name.

Every product in this category is a button with no inside: you press Enhance and get a picture, and if you disagree with it your only options are to accept it or throw it away. This tool reports the measurement behind every decision — how many stops the exposure was out, what fraction of the picture was clipped, what the noise floor measured — and turns each one into a switch. A correction you can see the reasoning for is one you can argue with, and one you learn from.

A batch feature that applies one setting to twenty photographs is not automatic enhancement; it is a preset with extra steps. Here the scene and the strength are shared, but every photograph is measured on its own and gets its own numbers — so the shots taken indoors have their cast removed while the ones taken at noon do not. Anything you adjust by hand applies across the batch on top of each photograph's own decision, and the finished set downloads as one archive.

A phone photograph is very often stored rotated, with a tag recording which way up it should be shown. A tool that reads the stored pixels instead of the tag analyses and corrects a picture the user has never seen — and then exports it sideways. The orientation is applied here before anything is measured, so the histogram, the cast estimate and the focus reading all describe the photograph as you see it.

Use it when you want the photograph you took, only right — and when you would rather not work out which of six things is wrong with it first. If you already know the look you want and intend to build it, the filters tool is the better starting point.

Common jobs

• A folder of holiday photos. Each one is measured and corrected from its own histogram, so the ones shot indoors get their cast removed and the ones shot at noon do not. The whole batch downloads as a single archive.
• Photographs shot in low light. The Low light scene lifts shadows harder, denoises harder and sharpens softer — three coupled moves, because lifting shadows is exactly what makes sensor noise visible.
• Portraits. The Portrait scene protects skin: clarity is reduced almost to nothing, because large-radius local contrast on a face finds every pore, and the shadow lift is raised because underexposed faces are the most common portrait fault.
• Scanned documents and receipts. The Scanned scene drives paper to full white and text to full black, removes saturation and pushes local contrast hard — the opposite of what a photograph wants on every axis.
• Faded prints and old scans. The Faded scene uses per-channel levels, the only mode that can pull three differently-faded channels back independently, and removes the yellow cast that age leaves behind.

Every figure below was measured on the reference machine with the accelerated path forced, not estimated. Where a device has no hardware acceleration the same code runs in software and produces the same picture, more slowly.

Numbers

• A 12-megapixel photograph is fully measured in about 40 milliseconds.
• The noise estimator recovers a known noise level to within one percent.
• The haze estimator recovers a known transmission to within three parts in a hundred.
• The accelerated and software paths agree to within one level out of 255 on every operation.
• Correcting a 12-megapixel photograph takes well under a second with hardware acceleration.
• A preview updates while a slider is still moving, and is never encoded to a file.

A 12-megapixel photograph is measured in about 40 milliseconds and corrected in well under a second with hardware acceleration. It all runs on your device, not our servers.

Automatic does not mean unattended. Three habits make a noticeable difference, and all three are about giving the measurement better material to work with.

Do this

• Crop before you enhance. Every measurement is taken over the whole frame, so a large expanse of sky or a black border pulls the exposure and levels decisions toward itself.
• Start from the original file, not a screenshot or a re-saved copy. Compression artefacts read as noise, and the denoising decision will spend its budget on them.
• Pick the scene that matches what you shot before adjusting anything by hand. A scene changes the targets the whole correction aims at, so it is a larger lever than any single slider.
• Turn Auto strength down rather than switching stages off, if the result is simply too much. It scales the entire decision back toward the original in one control.
• Use the before-and-after comparison rather than judging from memory. Corrections that look strong in isolation are usually right, and ones that look right in isolation are usually not enough.

When not to use it

• When you want a specific look rather than a correction — reach for the filters tool, which has curves, split toning and imported presets.
• When the photograph is heavily damaged by noise; this tool denoises what it measures, and a dedicated pass in the noise reduction tool goes further.
• When the picture needs to be larger rather than better; the image upscaler enlarges with an on-device model instead of interpolation.
• When you are working to a fixed file size; compress the image first, then enhance, so the correction is not undone by the encoder.
• For medical, forensic or scientific images, where any automatic tonal change may remove exactly the information you are looking at.

The photograph is decoded, measured, corrected and written on your device, not our servers, which is why there is no upload progress bar, no processing queue and no credit meter — and why a folder of family photographs stays a folder of family photographs. Where your device has hardware acceleration the tool uses it; where it does not, the same code runs more slowly and produces the same picture.

• Build a look by hand with curves, split toning and imported presets.

• Heavier denoising for photographs shot at high ISO.

• Fine-detail sharpening with halo and threshold control.

• Enlarge a photo with an on-device model instead of interpolation.

• Reduce a file below a size limit before enhancing it.

• Reframe and straighten before the correction runs.