GPT-image-2 vs DALL·E 3 vs Midjourney v7: 50-Prompt Benchmark

Benchmark setup: 50 standardized prompts across 5 categories — photoreal portraits, product staging, illustration, architecture, and abstract. Each prompt run identically through GPT-image-2 (medium quality), DALL·E 3, and Midjourney v7. Scored by 3 raters, 1–5 scale, on prompt fidelity and aesthetic quality. Full prompt list and outputs at the bottom.

Headline numbers

• Prompt fidelity: GPT-image-2 4.4 / DALL·E 3 3.6 / Midjourney v7 3.2.
• Aesthetic quality: GPT-image-2 4.1 / DALL·E 3 3.7 / Midjourney v7 4.5.
• Combined: GPT-image-2 4.25 / Midjourney v7 3.85 / DALL·E 3 3.65.

By category

Photoreal portraits

GPT-image-2 wins on prompt adherence — the right age, expression, and clothing. Midjourney wins on aesthetics; portraits look more 'magazine'. DALL·E 3 lagged on both.

Product staging

GPT-image-2 dominates here. Both DALL·E 3 and Midjourney often invent product features when given a written description; GPT-image-2 was strict.

Illustration

Midjourney v7 wins this category by a wide margin — its house style is its strength.

Architecture

GPT-image-2 wins on geometric consistency. The other two drift on perspective and proportions in long prompts.

Abstract

Roughly tied. All three produced striking images. Pick by personal aesthetic.

Cost-adjusted

Per-image cost matters. Midjourney is roughly $0.06 per image on a Standard plan; DALL·E 3 ~$0.04 (or 'free' inside ChatGPT Plus); GPT-image-2 medium ~$0.05. They're all close enough that quality should drive the choice, not price.

What we'd actually use

• Editorial / branded illustration: Midjourney v7.
• Product mockups, advertising stills, anything where prompt fidelity matters: GPT-image-2.
• Casual ChatGPT users not paying separate API bills: DALL·E 3 stays useful.