A 6×6 camera-angle reference sheet as a JSON prompt: thirty-six numbered panels, each with a Vietnamese caption naming the angle. One image.
One run, so there’s no holds and varies here. But one grid is thirty-six instructions, and it can be checked panel by panel.
The grid held completely. Thirty-six panels, numbered 1 to 36 in the top-left corner, in order, each with its caption in Vietnamese with the diacritics intact: Chính diện toàn thân, Góc nghiêng Hà Lan, Top-down nằm ngửa. The overhead panels (15–18) really are shot from above, which the eight-panel sheet couldn’t manage for its top-down view. Thirty-six enumerated entries, and the enumeration held again.
Needs a reference image, and the reference beat the text. The wardrobe line says “exact match to the reference”, and the result is a red-and-blue superhero bodysuit in every panel. But the panel list was clearly written for a different outfit: panel 9 asks for “the back of the corduroy shirt”, 23 and 25 for a “knit sweater”, 24 for “thick-frame brown glasses”. The picture overruled all of those leftovers. Panel 25’s fabric macro shows the suit’s webbing instead of corduroy, and panel 5’s “hands resting in pockets” has no pockets to use.
The glasses appear only in the two panels built around them: 24 (adjusting them) and 31 (macro of the eyes and frame). Panels 3 and 8 also mention glasses in passing, and both came back without them. A prop that is the subject of a panel gets drawn. A prop mentioned as a detail gets dropped.
