Why AI Upscaling Changes Image Structure: Sharpness vs Fidelity

The constraint being tested

The experiment ran on August 15, 2026 using an editing endpoint then labeled Qwen Image 2 Pro by a third-party provider. The request specified 1152×2048 PNG output. That service label is not presented as an independently verified official checkpoint identity.

The edit was meant to preserve crop, pose, silhouette, occlusion, lighting and uncertainty, without filling in unseen structure. Each result was Lanczos-downsampled back to the original 480×852 grid for metric comparison.

Samples A–E below are anonymous labels created for this article. Original images and asset identifiers are not published; measurements and visual observations come from the contemporary experiment report.

High similarity did not mean an accepted output

Five completed outputs covered local samples from two source scenes; one additional request did not complete. A failed request is separate from a failed visual review, and this small sample cannot establish a model-wide failure rate.

What the measurements can and cannot tell you

NCC is normalized cross-correlation. In this downsampled comparison it mainly indicated similar low-frequency composition. It could not verify that newly synthesized details were true. SSIM was another supporting signal, not a guarantee of unchanged content.

The detail ratio was a high-frequency comparison from this audit script, not a universal quality score. It may increase with useful sharpening or invented texture. We do not derive a transferable pass threshold from it.

Blur and occlusion carry information: they mark details that the input cannot establish. A plausible visual interpretation is not the same as recovering what was actually present.

Why keep a Lanczos baseline?

Deterministic interpolation cannot recover missing detail, but it provides a baseline that preserves the source structure. If an edit changes a silhouette, contact point or shadow boundary, describing it only as an upscale hides an important transformation.

That matters when the edited still becomes a video reference. A sharper image with a different starting pose may lead the next model into the wrong continuation. Whether to use the edit depends on the fidelity required by the next task.

A review order you can reuse

Define what must stay fixed before judging sharpness. Otherwise the immediate impression of clarity can distract from geometric changes.

1. Keep the original and a deterministic interpolation control, aligned for comparison.
2. Inspect crop, silhouette, cast size, occlusion and object contact points separately.
3. Treat blurred source regions as uncertain; do not treat generated details as original evidence.
4. Record request failure, completed-but-rejected output and pending review as different states.
5. Use your own publishable example assets for public comparisons, with metadata reviewed separately.

When generative refinement still makes sense

We rejected these outputs as high-fidelity production references. That does not mean generative editing is useless: a concept image that permits redesign and a reference that must remain faithful serve different purposes.

Creative redraws can be judged for visual appeal. Identity, motion or evidential fidelity require structural checks first. This article documents a failure analysis and does not establish a general image-quality ranking.