Prompt Format Probe

Azure gpt-image-2 · gpt-5.4-mini · quality=low · 1536×1024 · 2 repeats each · 2026-05-27
Same scene (family beach walk at dawn), 8 different structural formats for the user message.
style_score = fraction of watercolor style markers in revised_prompt (max 11 markers).
comp_score = fraction of composition markers (max 7 markers).

FormatImagesScores
A — Prose paragraph
925ch input
A — Prose paragraph
33.1s · 2555KB
A — Prose paragraph
44.1s · 2568KB
style 91%
comp 100%
38.6s avg
rp 902ch
B — JSON object
1058ch input
B — JSON object
123.3s · 2722KB
B — JSON object
64.3s · 2967KB
style 91%
comp 100%
93.8s avg
rp 928ch
C — XML tags
1050ch input
C — XML tags
112.3s · 2639KB
C — XML tags
104.5s · 2655KB
style 91%
comp 100%
108.4s avg
rp 899ch
D — Markdown sections
1019ch input
D — Markdown sections
94.8s · 2688KB
D — Markdown sections
136.0s · 2724KB
style 91%
comp 100%
115.4s avg
rp 968ch
E — Tag list (SD/MJ style)
529ch input
E — Tag list (SD/MJ style)
117.9s · 3254KB
E — Tag list (SD/MJ style)
92.1s · 2561KB
style 82%
comp 100%
105.0s avg
rp 605ch
F — B_template (production format)
1743ch input
F — B_template (production format)
207.8s · 2843KB
F — B_template (production format)
97.8s · 2588KB
style 91%
comp 86%
152.8s avg
rp 840ch
G — YAML
1017ch input
G — YAML
111.7s · 2650KB
G — YAML
135.4s · 2673KB
style 91%
comp 100%
123.6s avg
rp 928ch
H — Terse single sentence
157ch input
H — Terse single sentence
69.6s · 2502KB
H — Terse single sentence
75.1s · 2578KB
style 31%
comp 63%
72.3s avg
rp 403ch