What Litesum Enhance can and cannot do

This page is deliberately blunt. Enlargement software is sold with a great deal of exaggeration, and you are better served by knowing where this one is strong and where it is not.

What enlargement actually is

When an image is made four times larger, fifteen out of every sixteen pixels in the result were never in your file. They have to come from somewhere. A plain resize averages the neighbours, which is safe and soft. An AI model has been trained on millions of images and produces what such a picture usually looks like at that size.

That is the whole trade. The AI result is often much better. It is also, always, an informed guess.

Reconstructed detail is plausible, not recovered. It may not match what was in front of the camera. Do not rely on it to identify a person, read a number plate, or settle a question of fact. If the information is not in your file, nothing here can bring it back — it can only produce something convincing.

Where it is genuinely good

Measured against a known-correct answer, on a benchmark you can reproduce:

• Diagonal lines and line art. Structural similarity 0.93 against 0.79 for a good conventional resize. Diagonals stay continuous instead of turning into staircases.
• Lettering and logos. 0.95 against 0.86. Letter edges stay crisp, and the shapes are not altered.
• Architecture, straight edges, hard boundaries. 0.95 against 0.88.
• Flat-colour illustration and rendered artwork. Clean outlines survive; flat areas stay flat.

It manages this while producing fewer haloes than a sharpened conventional resize — 0.2 % of pixels overshooting on line art against 7 %. It is reconstructing edges rather than just increasing contrast around them.

Where it is not better, and we would rather tell you

On our benchmark the AI result is no better than a good conventional resize, and by some measures worse, on:

• Dense foliage and fur. Leaves and strands can merge into clumps. The AI keeps them more separate than a plain resize does, but both lose the finest detail, and the AI moves the image further from the true answer by one common measure while looking better by another.
• Smooth gradients and clear skies. The model adds a very faint texture that should not be there. On these areas we now hand the result back to a high-quality conventional resize entirely, which is the right answer — there is no detail in a smooth ramp to reconstruct — but it does mean the AI is contributing nothing there.
• Interface screenshots with hairline rules. One-pixel lines that have been reduced are difficult for anything to bring back.

If your work is mostly photographs of natural texture, try Litesum Enhance on your own images before committing to it. This page exists so that decision is made on what the tool does rather than on what we claim for it.

What photographs run, and why it changed

We tested pretrained AI models against a high-quality conventional enlargement on real photographs — skin, fur, snow, sky, old film scans — and the conventional method won on eleven of twelve, by both of the standard fidelity measures. Those models tend to reinterpret fine, chaotic texture — grain, stubble, fur — where the measurement says that texture was real detail. So the conventional method became the photographic default.

Then we trained our own photographic network with that lesson built in: purely to be faithful to the original, with nothing in its training that rewards invented texture. On the same twelve photographs it beat the conventional method on all twelve, by both measures. So at 4× it is now the photographic default — moved by the same rule that made classical the default before it — and every result is still checked region by region against a faithful resize. At 2× photographs keep the conventional method: we have not trained or measured a 2× network, and defaults follow measurements here.

The GAN-style network remains available for photographs as Generative Photo, clearly labelled: it invents plausible texture, the result can look subjectively richer, and what it adds may not have been in your original. That trade is yours to make, knowingly, not ours to make for you silently — and it is now a dial, not a switch. The Detail synthesis slider sets how much of the invented texture to keep: at zero the job runs the faithful classical enlargement and no AI touches your image; at full you get everything the network adds that survives the region-by-region fidelity check; in between, the result is smoothly pulled toward the faithful version by at least the amount you dialled away, and the "how much the AI contributed" message reflects the position you chose.

We tell you when the AI did not do much

Every enlargement is checked against a faithful conventional resize of your original, region by region. Where the AI has not improved on it, that region is handed back — and if enough of the image comes back that way, the status line says so:

Some areas were kept as a faithful resize, where the AI could not improve on the original.

or, when almost all of it does:

The AI found little here to reconstruct, so most of this result is a high-quality resize of your original.

Nothing is said when the AI did the work, so the message means something when it appears. The same figure is written into the batch report, so an unattended queue tells you the same thing.

We would rather say this than let you pay for an AI enlargement and quietly receive a resize. It also tells you something useful: an image that reports little AI contribution is one where the enlargement is about as good as it is going to get, and no amount of changing the settings will alter that.

And we tell you before you start, not only afterwards

On a large image, a 4× enlargement can take minutes. Waiting all of that to be told the AI had little to add is not much use, so the same check is run in advance, on a small sample taken from across your picture. If it expects most of the result to be a faithful resize, the panel says so under What to expect before you press Enhance:

Expected: much of this image will come back as a faithful resize rather than AI detail. Estimated from a sample, so the finished result may differ.

It is an estimate and it is described as one. Tested across our benchmark scenes it landed within 15 percentage points of the finished result on every one of them, at both 2× and 4×, and it is usually much closer than that — but it looks at a small part of your image, not all of it, so the finished figure is the one to trust. When it is wrong it more often expects too little than too much.

Nothing appears when the AI is expected to do the work, and nothing appears on small images, where the enlargement itself is quicker than the estimate would be.

Hard limits

Output size	Original size, 2×, 4×, any percentage from 1% to 400% — reduction included — and an exact width or height. 3× and 6× as trained model scales, a target megapixel count, a print-size calculator and pixel density in PPI are planned, not present.
Crop	Free, or one of six fixed shapes (1:1, 3:2, 2:3, 4:3, 3:4, 16:9), applied before processing and never written to your original. Undo, redo and a return to the whole photograph. Protection survives a crop being changed: it is remembered against the whole photograph, so tightening keeps it on what is still inside and widening brings back what was hidden. Changing a crop clears the brush's Undo history — its steps refer to the previous crop's pixels — and Export waits until the new crop has been enhanced. Rotating and flipping keep them as well, and turn them with the picture. Straightening by angle, custom aspect ratios and any kind of generative fill or extension are not here — Litesum Enhance chooses what the output is of; it does not invent picture outside your photograph.
Maximum output	32,768 pixels on a side, 500 megapixels total, and never more than the memory you actually have free. Litesum Enhance tells you before it starts, and suggests a scale that will work.
Colour depth	8 bits per channel throughout. 16-bit files can be opened; the depth is reduced. Full 16-bit is a planned milestone.
Formats in and out	JPEG, PNG, WebP, TIFF. No HEIC, AVIF, BMP or camera RAW yet.
Faces	Face recovery automatically applies recovery to eligible faces — small, soft ones in the finished image. Eligible means the size rule admits it, which is not the same as knowing it will look better: the size rule is a safeguard, not a quality judgement. It is not a general improvement to portraits: a face larger than 512 pixels in the output gets progressively less of it, and one half again as large gets none, because past that point the model removes detail rather than adding it. Some of a rebuilt face is the model's work rather than your photograph's -- moles, fine lines and skin texture go first -- so compare before exporting a portrait. It has been tested on 32 photographs of people, not on a large or representative set. When you reduce a photograph, faces are judged at the size they are in your original, not at the smaller size you asked for — so a 600-pixel face reduced to 300 is treated as a 600-pixel face and gets little or no recovery.
Operating system	Windows only. macOS and Linux builds compile but have never been run.
Text	Very small text may not survive enlargement legibly. Letters are enlarged, not read — nothing here understands what the text says, so it cannot guarantee the shapes remain the same characters.

Things that cannot be recovered by anything

Detail finer than the pixels that remain. A woven fabric with a five-pixel pattern, reduced to a quarter size, is gone. Both the AI and a conventional resize score essentially zero on recovering it. The AI will invent a weave; it will not be your weave.

Severe motion blur or focus error. There is no deblurring in this version.

Being honest about the benchmark

The benchmark that produced the figures above uses ten scenes drawn in code, not photographs. That was chosen so there is a known-correct answer to measure against, which no collected photograph can give — but it means these numbers say nothing about sensor noise, lens character or real skin. Adding real photographs is on the list.

Every figure quoted here, including the unflattering ones, is in docs/benchmark.md in the source repository, along with the images.