Midjourney thinking mode tested: loses to GPT Image 8.0 vs 5.3 on complex prompt
koltregaskes · x · 2026-10-12
Midjourney is experimenting with a "thinking model" to close the gap with GPT-Image on prompt adherence, now live on its alpha site.
@koltregaskes ran a quick head-to-head using the same elaborate prompt (Times Square-style NYC intersection at dusk, 150+ pedestrians, heavy signage text, Canon EOS R5 camera metadata):
- GPT Image (Image 1): 8.0/10 — excellent composition, crowd density, and accurate text, with only minor omissions
- Midjourney thinking mode (Image 2): 5.3/10 — strong atmosphere and lighting, but too few pedestrians and several text errors
- Midjourney without thinking (Image 3): 5.0/10 — good environmental detail but weaker ad accuracy and sparse crowds
Verdict: GPT Image wins comfortably; the two Midjourney versions are close, with thinking mode slightly ahead on text accuracy. The author notes that ignoring accuracy, Midjourney's output simply looks nicer.
More from Multimodal
- Seedance 2 video generation demo: 'Torneo de Mujeres 2' — Ok-Nerve941 · 2026-10-12
- Brand designer tests Midjourney V8.2 and Opus 5.5 on a full fake beverage brand — gizakdag · 2026-10-12
- Melty MCP generates music mashups from a single prompt — julianweisser · 2026-10-12
- Stunning AI image of a tiny mermaid in a glass of water — LudovicCreator · 2026-10-12
- Free anime character & prompt library launched for AI art creators — SenpaizCreations · 2026-10-12
- AI-made spy thriller short film 'The Countess and the Spy' showcases cinematic quality — SuspiciousPrune4 · 2026-10-12