Prompt engineering for AI image generation has accumulated more folklore than evidence. We ran a controlled experiment in May 2026: 60 prompts in five subject categories, run twice on each of Midjourney v7, Flux 2 Pro and Imagen 4 — once with the 'rule' applied, once without — and scored on a blind 1-10 quality rubric.

Four rules survived the test. Three did not.

Rules that consistently work

**1. Lead with the subject, end with the modifiers.** Putting the noun first and stacking style, lighting and camera details after improved scores by 18% on average across all three models. Models tokenize left-to-right and weight early tokens more heavily.

**2. Specify the lens or focal length.** Adding '85mm portrait lens', '24mm wide angle' or 'macro 100mm' improved composition quality 12-22% depending on subject. The numbers do not matter to the model — but they encode camera-style framing trained into the data.

**3. One style reference per generation.** 'Oil painting in the style of Rembrandt with watercolor accents and pixel art shading' produces mush. Picking one dominant style and adding texture words ('soft', 'crisp', 'painterly') beats mixing three style references.

**4. Negative prompts only on Stable Diffusion and Flux.** Midjourney's --no parameter works well. GPT Image-2 and Imagen 4 generally ignore explicit negation and respond better to positive rephrasing — 'serene' instead of 'no chaos'.

Rules that made no difference

**1. Adding artist names.** With most models filtering or remixing artist signatures, naming specific artists no longer moves output quality reliably. Style descriptors ('impressionist', 'art nouveau') work better.

**2. Magic suffix words.** '4k', 'masterpiece', 'best quality' once boosted Stable Diffusion 1.5 output. On modern models these are training-set noise and add nothing.

**3. Camera body model.** 'Shot on Canon EOS R5' vs 'shot on Sony A7R IV' produced statistically identical results. The lens, lighting and subject distance matter — the body name does not.

A starter template

For portraits: [subject], [pose and expression], [lighting], [lens], [style descriptor], [mood/atmosphere].

Example: A weathered fisherman, gazing toward stormy horizon, side-lit by golden hour sun, 85mm portrait lens, painterly oil-style finish, melancholic and reflective.

Browse our prompts library for 50+ tested examples across portrait, logo and anime categories.