How to write AI prompts for realistic people

Why people are the hardest subject

Viewers are experts at faces and bodies. We notice a slightly wrong hand, an impossible posture or skin that looks like plastic long before we notice a flaw in a building or a landscape. That makes people the subject where vague prompts fail most visibly. The fix is the same as for any photographic prompt, only stricter: describe what a camera would actually record, and leave as little as possible to the model's idea of an average person.

Invent the person, do not borrow one

Start from a person who does not exist. Describe an age range, a build, hair, clothing and an activity, and let the details be ordinary. Never prompt for the likeness of a real private individual, and never generate a real public figure in a situation that could be taken as genuine. Beyond the ethics, many jurisdictions recognise personality rights, which let people control the commercial use of their name, image and likeness. The rules differ by country, and an invented person avoids the question entirely.

Describe an activity, not a pose

People photographed doing something look natural; people told to pose look like stock images. Instead of a woman smiling at the camera, try a woman in her forties reading a paper map at a train station, holding a coffee in her other hand. An activity gives the body a reason for its posture, gives the hands something to hold and gives the face an expression that comes from the task rather than from the prompt.

Posture and weight

Real bodies carry weight. Someone leaning on a railing, sitting on the edge of a chair, shifting weight onto one leg while waiting: these are small physical facts that make a figure believable. Stiff, symmetrical, front-facing figures are one of the most common signs of a generated image. A single sentence about where the weight sits and what the person is leaning against often does more than several sentences about their appearance.

Clothing as material and fit

Describe clothes the way you would describe any other surface: a creased linen shirt with rolled sleeves, a heavy wool coat slightly too large, worn canvas trainers. Fabric that folds, stretches and wrinkles is one of the strongest realism cues, because it responds to the body underneath. Generic words such as stylish or elegant say nothing about material and usually produce clothes that look freshly unpacked.

Skin, without the retouching

Generated faces often come out smooth and even, like heavily retouched advertising. If you want a natural result, say so in physical terms: natural skin texture, visible pores, fine lines around the eyes, a light sunburn across the nose, no makeup. Describing the light helps too, because soft side light reveals texture gently while flat frontal light erases it. Avoid piling on terms such as perfect or flawless, which push the result further toward the plastic look.

Hands need a job

Hands are a well-known weak point of image generators. The most reliable approach is to give them something specific to do, holding a cup, resting on a table, tucked into a coat pocket, rather than leaving them free in the frame. Fewer visible fingers means fewer chances for an error. Always inspect hands before using an image: count the fingers, check how they grip objects, and look at where they meet the wrist.

Light on faces

The classic photographic setup, described in the three-point lighting tradition, uses a main light on the subject, a secondary light to soften shadows, and a light from behind to separate the subject from the background. You do not need the jargon to use the idea. Soft window light from one side, a gentle reflection filling the shadow side and a bright background behind the head describes the same arrangement in plain words. Overhead midday sun, by contrast, puts dark shadows under the eyes and is rarely flattering.

Framing and distance

Say how close the camera is. A head-and-shoulders portrait, a figure seen from a few metres away in a street, a small figure in a wide landscape: each one sets the scale at which faces and hands have to be rendered. Wider framings hide small errors and tell more of a story; close-ups expose everything. For the most natural results, a medium framing with a clear activity and a real setting is the most forgiving starting point.

Diversity by intention, not by accident

Image generators reflect patterns in their training data, and one study of Midjourney found that even neutral prompts returned unequal results by gender, skin colour and location. If you want a specific person, describe them specifically. If you are making a series, vary age, build and appearance deliberately rather than accepting whatever the model returns first, and review the set as a whole before you publish it.

Consent and synthetic media

The Partnership on AI's framework for responsible synthetic media is built around three ideas: consent, disclosure and transparency. For generated people, consent mostly means not using anyone's likeness without permission, including through reference photos. If you work from a photograph of a real person, for instance to keep a consistent character, make sure you have that person's clear agreement for the specific use, and never create images that sexualise, humiliate or impersonate anyone.

Disclosure rules to know

In the European Union, the AI Act defines a deep fake as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to be authentic. Deployers who generate such content must disclose that it was artificially generated or manipulated, with more limited disclosure for evidently artistic, creative, satirical or fictional work. The European Commission states that these transparency rules apply from August 2026. A realistic invented person in a context that could be taken as real deserves a label everywhere.

Label the file, not only the caption

A caption can be cropped away or lost when an image is reshared. Metadata standards exist to carry the information with the file: the IPTC Digital Source Type vocabulary includes values for media created using generative AI, and many tools write such information automatically. Keep it intact when you edit and export, and add a visible note wherever people could reasonably mistake the image for a photograph of a real person.

Where LarpGPT fits

LarpGPT is a library of photo prompts for travel, architecture, cars and lifestyle scenes, written around subject, composition, lighting and lens, and many of its lifestyle scenes can be adapted to include an invented person using the approach above. You copy a prompt and paste it into the image generator you already use. LarpGPT does not generate images; respecting other people's rights and disclosing synthetic content stay with you.

Further reading

• Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 3 and 50
• European Commission: AI Act, regulatory framework and application timeline
• Partnership on AI: Responsible Practices for Synthetic Media
• Personality rights (Wikipedia)
• Midjourney (Wikipedia), section on bias in generated images
• IPTC Digital Source Type NewsCodes