How to Write AI Photo Prompts That Work (2026)

Most AI photos look wrong in the same five ways — waxy skin, dead eyes, lighting that obeys no sun, hands from another dimension, and that unmistakable "AI did this" sheen. The difference between those and the portraits all over your feed is rarely the app. It is the prompt. This guide breaks down the formula behind prompts that actually work in 2026: what to specify, what to leave alone, and the mistakes that mark a photo as AI-generated from across the room.

Why prompt quality matters more than the app

Every major image model in 2026 — Gemini, ChatGPT's image editor, and the rest — is capable of a photorealistic portrait. What separates a good diya-glow portrait from a waxy mannequin is specificity. Vague prompts ("make me look cool") force the model to invent dozens of decisions: what light, what lens, what mood, what background. It invents them badly, averaging every photo it has ever seen into grey mush. A good prompt removes decisions. It tells the model exactly what the scene is, so the model's considerable skill goes into rendering rather than guessing.

The second reason is likeness. The number-one complaint about AI portraits is "it doesn't look like me." That is almost never a model problem — it is a prompt problem. Prompts that say "keep my face, expression and pose exactly the same" preserve likeness dramatically better than prompts that just say "of me." Every portrait prompt on PromptRoz carries an explicit likeness lock for this reason.

The 7-part formula

After testing dozens of prompts across 2026's viral trends, the ones that consistently produce postable results all contain the same seven ingredients. Not always in this order, but always present:

1. The likeness lock (subject)

State what must not change before you state what should. "Keep my face, expression and pose exactly the same" is the single highest-leverage sentence in AI portraiture. For couples, lock both faces: the pre-wedding shoot prompt does this, which is why two separate selfies can become one believable frame.

2. Style and medium

Name the visual language in concrete terms: "cinematic portrait," "photorealistic," "vertical portrait composition." Avoid stacking five styles ("cinematic photorealistic oil-painting anime") — contradictory style words are how you get mush. One clear style, stated once.

3. Light direction

Light is 80% of photography and 80% of AI portraiture. Never write just "good lighting." Say where the light comes from and what colour it is: "warm golden light rising from diyas below the frame," "golden rim light on hair, soft glow on faces." If your result looks flat, the light description was vague. Fix the light first, everything else second.

4. Lens, framing and composition

"Vertical portrait," "close-cropped," "rule of thirds" — these are not decoration. They control the single most common failure in AI portraits: the model drifting into a weird crop with half a shoulder and a floating elbow. State the framing and the model obeys it.

5. Colour and mood

Give the palette, not the vibe. "Deep maroon and gold palette, warm golden colour grade" produces a coherent image; "festive vibes" produces confetti soup. Two to three named colours are enough. The 1980s retro prompt works because its palette — grainy film warmth — is specified, not implied.

6. Texture and detail cues

Photorealism lives in micro-detail: "film grain," "soft bokeh," "natural skin texture." The word "photorealistic" alone is weak — models hear it constantly. Pair it with one texture cue and the render tightens visibly. Bokeh deserves special mention: "dozens of blurred diyas creating golden bokeh" tells the model exactly what the background should dissolve into.

7. The negatives (what to exclude)

End every portrait prompt with exclusions: "no text, no watermark." AI models love sneaking garbled text and fake watermarks into images. One exclusion line at the end prevents the most common post-processing headache. For couple shots, add "simple natural hand positions" — it does not fix hands, but it reduces the ambitious poses where hands break.

Three worked examples

Watch the formula in the wild:

• The diya portrait. The Diwali diya-glow prompt locks the face first, then rebuilds only the light — warm from below, bokeh behind. It changes almost nothing about you, which is precisely why the likeness survives and the photo looks expensive.
• The saree trend. The Nano Banana saree prompt pairs a strong style cue ("90s Bollywood retro") with a film-grain texture cue. The grain does quiet work: it hides the micro-imperfections that scream AI.
• The retro photo. The 1980s retro prompt adds an identity-lock trick on top of the formula — it is the likeness-lock principle taken one step further, and it is why that trend's results stayed recognisable.

Five mistakes that mark a photo as AI

1. Uploading a bad source photo. The model cannot invent detail. A blurry, dark, far-away selfie becomes a blurry stranger. Sharp, well-lit, face-forward — every time.
2. Asking for too many changes at once. New outfit plus new background plus new pose plus new lighting in one prompt is how faces melt. Change the scene; keep the person.
3. Forgetting the light. If your portrait looks like a passport photo with a filter, the prompt never described light. Go back to part 3 of the formula.
4. Ignoring hands. Check every finger before posting, especially in couple and embrace shots. Regenerate rather than repair — a second run is faster than any fix.
5. Baking in text. Names, dates and wishes belong in your caption app, not in the image prompt. AI text has improved, but a misspelled name on a greeting card is forever. (Our greeting card prompt is the exception — it locks the exact wording, and even then, zoom in and check.)

Your 30-second checklist

Before you hit generate on any portrait prompt: sharp source photo? Face locked in the prompt? Light described with direction and colour? One style, one palette? Exclusions at the end? Five yeses, and you are already ahead of 90% of AI portraits on the feed. Browse the photo prompts hub or the seasonal Diwali collection to see the formula applied — every prompt there follows it, and every one is free to copy.

FAQQuick answers

Guide last reviewed 29 September 2026 · Free to use