Restore

Perception Beginner Runs on this server Real-time

The one lab with no neural network in it. Edge-aware filters amplify local contrast while leaving flat areas alone — a useful baseline for what 'AI' is and isn't.

Model: OpenCV detailEnhance
Parameters: None · no learned parameters — hand-written maths
Input: Image
Output: Image
Typical time: 0.1–0.5s
Last run: —

01 Your image

Drop, paste or choose JPG · PNG · WEBP · max 10MB

Samples

02 Detail

Detail strength 10

How large a neighbourhood counts as “detail”. Low values sharpen fine texture; high values start lifting broad shapes and look artificial.

Edge preservation 0.15

How different two pixels must be before the filter treats the boundary as an edge and stops smoothing across it.

03 Finishing

Sharpening 0.00

An unsharp mask. Push it and you will see halos appear along high-contrast edges — the classic over-sharpened look.

Contrast 1.00

“Overdone” is included on purpose. Seeing the failure mode is the fastest way to learn where the useful range ends.

Before / after Drag the divider to compare

Nothing enhanced yet

A slightly soft or low-contrast photograph shows the effect most clearly. An already-sharp image has little headroom.

What actually changed

Nothing was added. An edge-preserving filter split the image into a smooth base layer and a detail layer, multiplied the detail layer, and put them back together. Every step is an equation somebody wrote by hand in the 2000s — there are no learned parameters anywhere in this lab.

Why this is the useful baseline A generative upscaler faced with a blurred face will produce a sharp face — but the detail is invented, plausible rather than recovered, and it can invent a different person. This filter cannot do that. It can only redistribute contrast that is already present in the pixels. When you next see a dramatic “enhance” result, the question worth asking is which of the two you are looking at.

How it works

An edge-preserving filter separates the image into a smooth base layer and a detail layer, boosts the detail layer, then recombines them. Every step is a fixed equation written by a person. Nothing was learned from data.

Input Your image, unchanged

Split Edge-preserving filter separates base from detail

Amplify Detail layer scaled, then optional unsharp mask

Output Recombined image — no new information

Full identifier: cv2.detailEnhance — domain transform filter
Where it runs: On this server, on CPU
Parameters: None — this is not a learned model

Deterministic and fast, but strictly limited to redistributing contrast that already exists in the pixels.

What to notice

It cannot add information that isn't there. It makes existing detail more visible; it does not reconstruct lost detail.
Push it too far and you get halos around edges and amplified sensor noise — the classic over-sharpened look.
Compare this to what a generative upscaler does: that one invents plausible detail. This one doesn't invent anything.

A note on language

This page says “the model predicts”, not “the model knows”. That is deliberate. None of these systems understand the images or sentences you give them; they map inputs to outputs using patterns fixed at training time. The difference matters most exactly when the output is impressive.