Where the work actually stands.

Unreleased

Detail synthesis: invention is now a dial, not a switch. The Generative Photo image type gains a 0–100 % Detail synthesis slider. At 100 % it is exactly what Generative Photo produced before; at 0 % the job runs the faithful classical enlargement and no AI touches the image; in between, the result is smoothly blended toward the faithful version — using the same region-by-region fidelity machinery that already guards every AI result, so the blend costs nothing extra. The contribution message is honest about the position you chose: at 40 % detail synthesis the AI contributed at most 40 %. The slider appears only on Generative Photo, because on every other image type faithfulness is the promise, not a preference.

The AI models look after themselves. At every start the application checks its models, rebuilds the record from files already on your computer when it can — offline, instantly — and downloads only what no local file can supply, with progress in the status bar. The Repair button does the same: local rebuild first, download second. And if the models are ever genuinely absent, the photographic image types no longer disappear from the picker — they use the faithful classical method and need no models at all.

Photographs at 4× now run Litesum's own fidelity-trained AI network. The default followed the measurements, twice. First we benchmarked pretrained AI models against a high-quality conventional enlargement on twelve real photographs, and the conventional method won eleven — so it became the default. Then we trained our own photographic network purely for faithfulness, with nothing in its training that rewards invented texture, and on the same twelve photographs it beat the conventional method on all twelve, by both standard fidelity measures. So at 4× it is now the photographic default, still checked region by region against a faithful resize. At 2× photographs keep the faithful classical enlargement — no 2× network has been trained or measured, and defaults follow measurements. The AI keeps the categories where the benchmark shows it clearly winning (Illustration and CGI, Text and Graphics), and the generative look remains available as Generative Photo, an explicit opt-in that says what it is: plausible invented texture, not guaranteed original. Classical jobs still start faster and report honestly that no AI ran.

Time estimates before a job starts. Once your computer has run a few jobs, the panel says how long the next one is likely to take, from measured history on your machine — and says nothing rather than guessing when it has no honest basis.

0.1.0 — not yet released

This build is not ready for sale. It is complete enough to install, run and be judged, and it is being recorded honestly so the remaining decisions can be made on evidence. release-checklist.md has the current scores and the reasons.

What works

Installs as one file. A 294 MB installer containing the application, the .NET runtime, the graphics runtime and all three AI models. Nothing is downloaded afterwards. Verified by installing into a clean folder, disconnecting nothing, and enhancing an image on a computer with no development tools.

Processes entirely on your computer. Driven through import, 4× enhancement and export while every network socket was sampled 196 times: zero connections. An automated audit fails the build if anyone adds an analytics library, a crash reporter or a network call in the enhancement path.

Works for 14 days without signing in, so it can be tried before it is bought.

2× and 4× enlargement, with output dimensions exactly right including on odd-numbered source sizes.

Six image types — General Photo, High Quality Photo, Low Resolution, Compressed Image, Illustration and CGI, Text and Graphics — as processing profiles over two networks, with the sharing stated in About rather than hidden.

Automatic mode measures the image, chooses a type, sets the three strengths and explains why in a sentence. Every value can be overridden.

A safety limit that judges each job against the memory actually free and refuses what cannot finish, naming a scale that will work.

Batch processing with pause, skip, cancel, failure isolation, a memory pre-flight per image, a choice of output format defaulting to lossless, optional folder-structure preservation, and a report recording what was done.

JPEG, PNG, WebP and TIFF in and out, with ICC profiles preserved, EXIF stripped by default including GPS, and orientation handled so images are never rotated twice.

What is honestly not good enough yet

Photographs of natural texture. On the benchmark the model is no better than a good conventional resize on foliage, fur, smooth gradients and screenshots, and by one common measure worse. It is genuinely strong on line art, lettering, architecture and flat artwork. Whether v1.0 ships as a graphics-and-text upscaler is an open decision.

A faint regular pattern on smooth areas. Reduced to conventional-resize levels on most flat content; a clear gradient still shows more than it should.

Not code-signed. Windows warns about the publisher on every install.

Windows only. macOS and Linux builds compile and have never been run.

No face recovery, and nothing here has been tested on faces.

Fixed during development, worth recording

• The published build shipped without DirectML.dll, losing GPU acceleration silently — every customer would have had processing roughly twenty times slower.
• The installer produced an application that could not enhance anything on a clean machine, because the licence gate needed an identity only the Launcher creates. It survived two rounds of testing because the development machine was already activated.
• Repairing one damaged model de-registered the other two, which then became unreachable offline.
• The memory guard measured installed memory rather than free memory, so a job needing 8.7 GB was called safe on a machine with 7 GB free.
• Batch silently re-encoded every JPEG at quality 92, discarding detail the model had just reconstructed, while the same image exported on its own was lossless.
• "Apply current settings to all" applied only the enlargement.
• Moving a slider did not turn automatic mode off, so hand-tuned values were overwritten without warning when the next image was opened.
• The enlargement picker announced ScaleOption { Scale = Two, Label = 2× } to screen readers.
• Uninstalling with the application open removed the shortcut and the Add/Remove Programs entry and then failed to delete anything, stranding 311 MB with no way to remove it.

Each has a regression test.