The risk map: where an upscaler invents detail

Every result in the studio is checked against the same photo enlarged plainly, cell by cell, and the cells where the result carries far more edge detail than the plain version are lit. This page is the method, the numbers it uses, and what it cannot tell you.

What is compared

Two pictures of the same size: the result, and the source enlarged with Lanczos3 to that size — the control, which is bigger pixels and nothing else. Both are reduced to brightness and run through a Laplacian, which responds to edges and texture. The map divides the picture into 16-pixel cells and, in each, takes the mean absolute Laplacian of the result over that of the control.

A ratio near 1 means the cell is as busy as the source made it. Sharpening raises it along edges the source already had. A reconstruction network that draws texture where the control is flat — pores on smooth skin, bark on a blurred trunk, a pattern on plain fabric, a reshaped letter — raises it fastest, because the control’s energy there is small.

The numbers

Every figure below is read from the code that draws the map
What	Value	Meaning
Cell size	16 px	Blocky on purpose: a grid a person can scan and send the loupe to
Scoring starts	2.4×	Edge energy ratio, result over control, below which a cell scores zero
Score saturates	4.8×	Ratio at which a cell's score reaches 1
Hot cell	score above 0.5	Ratio above 3.6×
Medium verdict	3% of cells hot	Tour the spots and decide whether you believe them
High verdict	15% of cells hot	Check before using it for anything that has to be true
Near-silent cells	result energy under 3	Never scored: noise level, nothing to judge

The hot-spot tour

Up to six of the strongest cells, spread apart so the tour does not visit one leaf six times, are numbered on the stage. N moves the loupe to the next, Shift+N to the previous, and the review panel lists the same spots in words — their number, their ratio and where they sit in the frame — so the tour works with a screen reader as well as with a mouse.

Faces first

When the face detector on this device has found faces, the tour starts inside them: up to three spots from the face cells, then the rest of the picture. Invented skin and reshaped eyes are the failure a portrait cannot carry, so they are the first thing shown.

Reviewed, not enforced

The save line counts the spots you actually visited — “3 of 6 hot spots reviewed”. Nothing stops a save with spots unvisited; the count is there so the person deciding knows how much of the review was done.

What it cannot tell you

Whether a lit cell is wrong. A lit cell is texture the plain enlargement does not have; whether that leaf or pore was really there is something no measurement of these two pictures can know.
Whether a dark cell is right. A model can move an edge without adding energy — a letter redrawn as a different letter at the same contrast. Text-like photos get a separate caution for that reason.
Anything about colour. The map works on brightness; ESPCN works on brightness only and cannot shift colour, and Real-ESRGAN's colour is checked by eye in the split view.
That the result is verified. The map shows where to look.

Questions

What does a lit cell mean?

That the result carries at least 2.4 times the edge energy of the same photo enlarged plainly with Lanczos3, in that 16-pixel cell. It means "look here", not "this is wrong": the map cannot know whether a leaf or a pore was really there.

Why compare with a plain enlargement?

Because a plain Lanczos3 enlargement adds nothing: it is the source's own detail, bigger. Anything the result has above it was added by the pipeline — sharpening where an edge already existed, invention where the plain enlargement is flat.

Why does scoring start at 2.4×?

Because sharpening what exists lifts edge energy to about twice the control. Measured on a real portrait during development, Lanczos3 plus detail and ESPCN stayed under about 2.9× in their strongest cells, while Real-ESRGAN's grown foliage reached about 7×. Starting at 2.4× and saturating at 4.8× keeps honest sharpening mostly dark and lights growth.

Does a low risk verdict mean the result is accurate?

No. It means the added detail sits mostly on edges the source already had. A model can still change the shape of an edge it sharpened — a letter, an eyelid — which is why the loupe and the 100% view exist. The map is never called a verification.

Does the risk map send my photo anywhere?

No. It is arithmetic on two pictures already in this tab, the result and the plain enlargement, and it runs in a few hundred milliseconds on this device.