Enhance a photo with local AI

Most of what makes a phone photo disappointing is light and colour: it is too dark, too yellow, too flat. None of that needs a model. The same measured pass every page on this site runs corrects exposure, white balance, range, contrast and shadows from the picture’s own histogram and prints every change in stops and kelvin. That result is on screen before anything else and it can be saved as it is.

A model earns its place for the things arithmetic cannot do: drawing plausible detail into a picture that has too few pixels, and smoothing the blocks and smears of heavy JPEG compression on the way. Those are reconstructions, not measurements, so they are offered as a separate, optional second step, run on top of the classical result, and labelled wherever their output appears.

The 2× detail model runs on all three colour channels at four times the size, and its output is halved with an exact two-by-two average, so the factor printed is the factor you get. The fast model enlarges brightness three times and lets colour follow smoothly, which is how it stays a quarter of a megabyte. Each file is checked against its recorded SHA-256 fingerprint after download and before it runs; a file that does not match is discarded.

Press Try on the loupe region and the model runs on the small region under the loupe — usually one tile, a second or two — and shows its output beside the original at 1:1. That run also measures how fast this device is, and the whole-photo button updates its prediction with the measured time. If the whole photo would take longer than a minute, or produce more than sixteen megapixels, the panel offers a smaller copy instead and says how long the full size would have taken.

Cancel stops at the next tile. The classical result is never replaced: a model result sits beside it until you save, the save button changes to say it includes reconstructed detail, and Use the classical result drops it. Change any slider and the model result is discarded, because it no longer describes the picture on screen.

• A model enlarges with plausible detail; it does not recover the original scene. Text, number plates and faces can come out confidently wrong.
• No face restoration, no colourisation, no generative fill. The models here enlarge and clean; they do not repaint.
• Speed depends on your device. The time printed is measured on it, not promised.

Is my photo uploaded to run the AI?

No. The model file comes to your browser, not your photo to a server. The runtime and the model execute inside this tab; the picture goes from a canvas into the model and back without ever being part of a network request. The only thing downloaded is the model itself, from our model file server, and only after you press the button that names its size.

How big is the download?

The 2× detail model is 4.87 MB and the fast 3× model is 0.24 MB. The first model you run also fetches the runtime that executes it, 14.03 MB, from this site. Both are kept in your browser afterwards, so the second run downloads nothing, and the Local AI panel shows what is held and clears it with one button.

Why is it slower than an app?

A phone app sends your photo to a data centre full of graphics cards. This page runs the model on your own processor, single-threaded, tile by tile. The registry's figure for the 2× model is about 1.2 s per tile, and a tile covers about 13 thousand source pixels, so a whole 12-megapixel photo would take many minutes. That is why the panel measures the speed on your device, predicts the time before you start, runs the loupe region first in a second or two, and offers a smaller copy when the whole photo would take longer than a minute.

What does the model invent?

Fine texture. A super-resolution model has learned what detail usually looks like at a larger size and draws plausible detail where the small picture had none — hair, brick, leaves. It is not the original scene; it is a convincing guess. That is why its output is labelled reconstructed in the loupe, on the save button and in the receipt, and why the classical result stays one click away.

Does it fix faces?

No. There is no face-restoration model here, because none has yet been found that is both licensed for this use and small enough to download into a browser. Super-resolution can change fine facial texture, so check eyes and hair at 1:1 before you save a portrait.

• The correctorOne press, then ten controls if the press did not get you all the way there.
• Improve image qualityA quality report first: grain, JPEG quality, blocking, edge energy, clipping and resolution, each with its fix.
• Image clarity enhancerClarity, sharpening and grain checked at 1:1 in a loupe on the busiest patch before you save.
• Batch enhancementGive every file its own measured correction and collect successful outputs in one ZIP.