The detail check

Two measurements, taken on every photo in both the studio and the queue: how much real detail the source carries, and whether the result has more of it than a plain enlargement. They answer the question every upscaler is asked — can this make my photo higher resolution? — with a number instead of a promise.

1. How much real detail the source carries

Several 192-pixel windows are sampled and the sharpest is kept. That window is halved and put back. A photo sharp to its own pixel grid loses most of its edge energy when halved; one that was itself an enlargement, or was soft to begin with, loses almost none, because it never had detail at that level. The test walks down the ladder — 2×, 4×, 8× — until the halving starts to cost something.

The result is an effective detail edge in pixels: “1,200 px wide, carries about 400 px of detail”. Below an acuity of 0.75, the photo was soft before anything ran, and the queue row and the studio both say so.

Calibration

Measured on this site’s own fixtures: a picture sharp to its own grid retains 0.21 to 0.32 of its edge energy through a halving; one that is itself an enlargement retains about 1.02. The threshold, 0.8, sits in the order of magnitude between them. The build runs those fixtures on every release, so the threshold cannot drift away from the evidence.

2. Whether the result added any

The output’s edge energy is compared with the same source interpolated to the same size, on a 512-pixel window in the same place. At or above 1.12×, the engine added measurable detail; below it, the output is bigger pixels.

What each figure is, read from the code that measures it
Figure	Value	Used for
Source window	192 px	Sharpest window, halved and restored
Flat retention	0.8	At or above: no detail at that level
Ladder	2×, 4×, 8×	Levels tested, and as far as one window can claim
Soft source	acuity under 0.75	Row and studio say the photo was soft already
Output window	512 px	Where the result is compared with interpolation
Reconstruction gain	1.12×	At or above: the engine added detail

The honest factor

On the low-to-high-resolution page the check runs first and recommends a factor: the effective detail edge times the model’s own magnification (four for Real-ESRGAN), against the photo’s long edge. Past that, the extra pixels are interpolation of a guess, and the studio says so rather than offering 8× as though it meant something.

What it will not claim

That the effective detail edge is a resolution standard. It is an estimate from one window.
That a “yes” means the added detail is true to the scene. That is the risk map's question, and it cannot fully answer it either.
That an upscale fixes blur. Motion blur and focus blur are not undone; the check simply measures that the detail was never there.

Questions

What does “carries about 400 px of real detail” mean?

That the finest detail in the photo is what a 400-pixel-wide picture can hold, whatever its pixel count says. A 1,200-pixel photo that was enlarged from 400 pixels at some point, or that was soft when it was taken, measures that way.

How is that measured?

The sharpest 192-pixel window of the photo is halved and enlarged back. If almost nothing changes — 0.8 or more of its edge energy retained — the photo carried nothing at that level, and the test steps down: 2×, 4×, 8×. The level it stops at gives the effective detail edge.

What does “output adds detail: yes” mean?

That the result carries at least 1.12 times the edge energy of the same source interpolated to the same size, measured on a 512-pixel window of the output. Below that, the extra pixels are bigger pixels of the detail the source already had.

Does “yes” mean the added detail is true?

No. It means a model drew measurable detail; whether that detail matches the scene is what the risk map and the loupe are for. The detail check is never described as a resolution standard.

Why measure one window and not the whole photo?

Cost. Bounded windows keep the check in single-digit milliseconds on any size of photo, so it runs on every file in a 500-file queue. The price is that it is an estimate from one place, and it is labelled as one.