From brief to approved image

A scene assembles itself. You pick the cast, the world and the style, and only type what happens. The application then builds the complete instruction itself, deterministically and visibly.

Making a new image

1. Pick the project and the category.
2. Assemble the cast, choose the world and the style, and type the action.
3. Pick a recipe, or adjust the model by hand. The estimated cost is shown before you generate.
4. The automatic check runs immediately on every variant, with a tick or a cross per mandatory trait.
5. You choose a variant; it becomes version 1 of a new asset and goes up for review.

The assembled instruction can always be expanded and read, literally as it goes to the model, and is stored in full with the generation. That is the only way to explain a fault afterwards.

Iterating on feedback

You place feedback with a marker on the image itself: click the detail that has to change and type what has to change about it. The marker is stored as a coordinate and travels with the next change request. The application then assembles an edit instruction in which the source image goes along as input, the canon goes along again, and the rest of the image must explicitly stay unchanged. The comparison view overlays the old and the new version with a slider, so it is immediately visible whether more has changed than was asked for.

Series, films and re-checking

• Series — a list of subjects becomes a run of tasks sharing one style, seed strategy and format. The result is a contact sheet showing average brightness, dominant palette and line weight side by side, so a deviating image stands out without zooming in.
• Films — for every shot a keyframe is first made and approved as a still image. Only then does that image go to the video model as a start frame. That is the approval gate: video is only made once the picture is already right.
• Re-checking — after a canon change the check runs again over the affected assets and sorts them into three piles: still fine, borderline and failing.
• Model bake-off — the same brief goes to several model profiles at once and the results are judged blind, because a brand name colours the judgement. The application then works out the cost per usable image; that number is the real verdict.