A gallery of generated images has two different jobs: letting someone see the range of a model, and helping them decide which candidates deserve a closer look. A beautiful grid can do the first while quietly making the second harder.
This note comes from Whispart, a studio developing and curating GAN-based images for physical artworks. Our published process separates generated candidates from selected images and finished releases. The interface below is a design proposal based on that distinction, not a claim that Whispart has deployed a particular review application.
Preserve the comparison before adding a score
Begin with a stable contact sheet. Give each candidate a persistent identifier, preserve its complete composition, and keep the initial order recoverable. Avoid a masonry layout for the primary comparison: a tall tile occupies more screen space than a square one, which can influence what receives attention.
Equal-sized containers are useful, but equal-sized crops are a different decision. Use object-fit: contain for the review thumbnail so that the interface does not silently remove an edge or a figure. A detail crop can be a separate, clearly labeled view.
Let the reviewer toggle between three views:
| View | Question it supports | Limitation |
|---|---|---|
| Complete contact sheet | Is there meaningful variation across this batch? | Small details are difficult to judge. |
| Single candidate, complete image | Does the composition hold together without its neighbors? | One image cannot demonstrate the range of a system. |
| Detail at a stated zoom | Is that mark deliberate-looking ambiguity or a distracting artifact? | Enlarged detail cannot decide the whole composition. |
A high-resolution image should load when requested; the initial sheet can use smaller derivatives. Keep the same candidate identifier attached to both. A thumbnail and its full-size file must never drift apart because filenames were sorted differently.
Make “undecided” a real state
An unchecked image has not necessarily been rejected. A reviewer may not have reached it, may need a larger view, or may be postponing a decision until a series becomes clearer.
For a small tool, an illustrative record could look like this:
{
"candidate_id": "study-017",
"review_state": "hold",
"observation": "The pale area competes with the figure at thumbnail size.",
"next_check": "Inspect the complete image on its own.",
"reviewed_at": "2026-09-09T04:00:00Z"
}
This is fictional example data. Its useful feature is the separation of observation from action. “Bad image” does not tell the next reviewer what to inspect. A bounded visual observation does.
Keep unreviewed, hold, shortlisted, and excluded distinct. Store a change history if multiple people can alter a decision. Shortlisting should not automatically authorize printing or publication; those are later decisions with different evidence.
Keep the studio's preferences legible
Whispart's approach values emotional tension, ambiguity and images that continue to hold attention. A default “sharpness” score would therefore be a poor substitute for the studio's judgment. Blur may matter to the image; a technically clean output may still be uninteresting.
For another practice, graphic clarity might be essential. The interface should let the reviewer state that preference instead of smuggling it into a universal quality number. Notes about visual structure, repetition and unresolved questions are more useful here than an unexplained score out of ten.
The Whispart process journal provides the concrete workflow behind this proposal: a visual field becomes candidates, then human selection, writing and material proofing determine what can become a physical work. A contact sheet supports the selection stage. It cannot supply all of the later decisions.
Published by Whispart. This article was drafted and checked by an AI assistant using Whispart's public studio documentation. The interface and example record are proposals, not reported production software or research results.
Top comments (0)