The Verdict
OpenAI gpt-image-2 wins the locked oil-painting canon. Donato/Manchess aesthetic comes through natively, text rendering is in a separate weight class, and complex prompt fidelity is a half-step ahead.
Gemini Direct is the value champion. Gorgeous fantasy-illustration aesthetic, clean 16:9 aspect, ~4× cheaper. Right tool for thumbnails, calendar cards, social fragments.
Replicate nano-banana-pro is dropped. Same underlying model as Gemini Direct (gemini-3-pro-image-preview), but with rate-limit pain (429s after one request) and binary-handling friction in n8n. No quality gain, all proxy cost.
Prompt A · The Cursed Tome (EXHIBIT A)
Halfling RJ-avatar in forest-green coat, size-5 boots on the table beside a battered grimoire labeled TIME SUCKING SOFTWARE. Four light sources, multiple characters, multiple readable signs. Tests prompt fidelity + text rendering hard.
Replicate column = same model as Gemini Direct, routed through Replicate's proxy. Compare to see if the proxy adds anything. (Spoiler: it doesn't.)
Prompt A · Consistency Re-Run
Identical prompt, fresh seed. Tests whether the halfling RJ-avatar stays recognizably the same character across generations — critical for canon work.
Both engines kept the halfling recognizably the same person. Canon-safe in both.
Prompt B · The Heart Tribute
Elven silver-haired singer in spotlight + dwarven lute-player in side-lighting as “architecture, not spotlight.” Diverse fantasy audience. Tests character distinction, dramatic lighting hierarchy, oil-painting aesthetic.
Gemini gave both women lutes — missed the singer/architecture hierarchy. OpenAI nailed it: pillar of light on the singer, warm side-light on the dwarf, exactly as the prompt described.
Prompt C · The Flashmob in the Rafters
Worm’s-eye view, 8-10 clockwork owls in synchronized chorus line, “CONTRACT SENT”/“IT IS DONE” banners, hidden detail (one owl out-of-sync, dropped banner, startled purple salamander). Tests extreme prompt fidelity + text-in-image at scale.
Both got the chorus line + salamander + banners + Wobblehat in the corner. Gemini won the cleaner worm’s-eye angle. OpenAI won the dramatic chiaroscuro lighting and brassy detail.