Apr 23, 2026 · 7 min read
Prompt Engineering for Photoreal AI Images
Prompts that read like photographer notes get photoreal results. Here's the formula — and the words that move the needle.
If your AI images come out 'almost real' — soft eyes, plasticky skin, the wrong kind of bokeh — the cause is almost always prompt vocabulary. GPT-image-2 reads photographer terminology fluently. Ours don't.
The four-line formula
- Subject + action: 'A 30-year-old chef plating risotto.'
- Camera + lens + aperture: '85mm prime at f/1.4, shallow depth of field.'
- Light + time: 'Diffused window light, golden hour, warm key.'
- Mood + reference: 'Editorial portrait, James Nachtwey.' (Use deceased or well-known names cautiously — see content policy.)
Vocabulary that works
- Lens: 24mm wide, 50mm normal, 85mm portrait, 135mm long — these change perspective.
- Light: rim light, fill, hair light, key, soft box, hard light, golden hour, blue hour.
- Film: Portra 400, Kodachrome, Tri-X, Cinestill 800T — powerful style anchors.
- Process: shot on medium format, scanned 35mm, instax, polaroid.
- Skin: dewy, matte, lived-in, freckles visible — counters the 'plastic AI face' look.
Vocabulary to avoid
- 'Hyper-realistic', 'ultra-detailed', '8K' — these often produce uncanny artifacts.
- 'Beautiful', 'stunning', 'gorgeous' — vague boosters, no signal.
- Negative phrasing in plain English. Use 'not' sparingly — describe what you want instead.
A worked example
Editorial portrait of a 30-year-old chef plating risotto. 85mm at f/1.8, shallow depth of field. Diffused window light, golden hour, warm key. Skin lived-in with freckles visible. Background: open restaurant kitchen, slightly out of focus. Shot on Portra 400, scanned 35mm.
That's 47 words and produces images that pass for editorial work. Notice that 95% of the prompt is photography — the AI handles the rest.