✨ The Verdict Inverted
Round 1 said: OpenAI gpt-image-2 wins the locked oil-painting canon. Then we tried curating the prompts — one hero, 2-3 light sources (warm+cool+sometimes magical), 2-3 readable text elements, striking dynamic camera angles — and sent them back through Gemini Direct.
The results below speak for themselves. Gemini, with curated prompts, delivers the Donato/Manchess oil-painting canon at 4× lower cost, faster, with no moderation rejections. The bottleneck was never the engine. It was the kitchen-sink prompt style dragging Gemini toward flat illustration.
Production rule going forward: Gemini Direct + curated prompts. The prompt is the editing.
v2 · Curated · Gemini Direct CURRENT STANDARD
Each scene has one clear hero, 2-3 named visual elements, 2-3 readable text elements, a striking camera angle, and 2-3 light sources (always warm+cool, sometimes magical). The through-line phrase: "An inviting, mischievous, magical, medieval place where the outcasts belong."
A · The Cursed Tome
Camera: Low-hero angle, halfling leaning into the frame next to the tome. Lights: warm hearth, cool moonlight, magical green miasma. Text: TIME SUCKING SOFTWARE · EXHIBIT A.
Show the original v1 bake-off (engine comparison · superseded)
v1 · Original Bake-Off (for reference)
What we tested: 3 maximalist prompts (~600 words each, 14 named visual elements, 4 light sources, 7 readable text elements) through three engines: Gemini Direct, OpenAI gpt-image-2, Replicate nano-banana-pro.
What we concluded at the time: OpenAI wins on canon adherence, Gemini wins on cost. Replicate was dropped — same model as Gemini Direct under the hood, but with rate-limit pain and proxy friction.
What we missed: The kitchen-sink prompts were pushing Gemini toward flat illustration. The engine was capable of oil-painting realism the whole time — we just hadn't given it room to breathe.










