● tokenstore

Writing Prompts for Explicit AI Images

The anatomy of a strong NSFW image prompt — subject, pose, framing, lighting, and style — with concrete-vs-vague rewrites and worked examples for Nureta's uncensored image model.

3 min read

The single biggest lever on an NSFW image is the prompt. The seahorse-image model is uncensored and anatomy-capable, so it will render whatever you actually describe — which means a vague prompt gets you a vague image and a precise prompt gets you a precise one. The skill is not coaxing the model past a filter; there is no filter. The skill is being specific.

This guide breaks a prompt into five parts you can fill in every time, shows the difference between concrete and vague wording, and covers how to steer away from things you do not want. Everything applies whether you prompt in the playground or through the image API.

The five parts of a prompt

A reliable explicit prompt names five things in order. You do not need flowery language — you need each of these present:

  • Subject — who is in frame: body type, age-of-adult descriptors, hair, skin, expression.
  • Pose or act — what they are doing: the position, the contact, where hands and hips are.
  • Framing — the shot: full-body, close-up, low-angle, three-quarter view, POV.
  • Lighting — the source and mood: warm lamplight, hard flash, soft window light, backlit.
  • Style — the render: photographic, film grain, editorial, matte, glossy, particular lens feel.
Order matters less than presence. If all five parts are in the prompt somewhere, the model has enough to render a specific image. Drop pose and framing and it falls back to a generic centered portrait.

Concrete beats vague

Vague words make the model guess, and its guess is the statistical average — soft, centered, unremarkable. Concrete words remove the guessing. Compare:

  • Vague: a sexy woman. Concrete: a nude woman lying on her back on rumpled sheets, one knee raised, seen from a low angle at the foot of the bed.
  • Vague: two people having sex. Concrete: a couple in missionary position, her legs around his waist, close side-on shot at hip height, warm lamplight.
  • Vague: good lighting. Concrete: single hard flash from camera-left, deep shadows, glossy skin.

Notice the concrete versions never get more explicit than the vague ones — they get more specific. That is the whole game.

Framing and camera language

Framing is where most prompts leave value on the table. The model responds to real camera terms: close-up, full-body, low-angle, high-angle, three-quarter view, over-the-shoulder, POV, wide shot. State the distance and the angle explicitly. close, low-angle shot renders a very different image than wide, high-angle shot from the same subject and act — and if you say nothing, you get a flat eye-level average.

Steering away from what you don't want

There is no separate negative-prompt field, so steer with positive description instead. Do not write not clothed — write fully nude. Do not write no harsh light — write soft diffused light. If the model keeps adding something you dislike, the fix is to describe the thing you want in its place. Positive, concrete phrasing overrides an unwanted default far more reliably than a negation the model may read as a keyword to include.

A worked example

Start rough and layer the five parts in. a woman on a bed becomes a nude woman on a bed, then a nude woman kneeling on a bed, arched back, then ... three-quarter view from behind, then ... warm window light from the left, late afternoon, then ... shot on 35mm film, fine grain. Each pass adds one dimension. Here is the finished prompt going to the API:

curl -s -X POST "$TOKENSTORE_URL/api/v3/images/generations/tasks" \
  -H "Authorization: Bearer sk-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seahorse-image",
    "content": [
      { "type": "text", "text": "a nude woman kneeling on a bed, arched back, three-quarter view from behind, warm window light from the left, late afternoon, shot on 35mm film, fine grain" }
    ],
    "size": "2K"
  }'

At $0.05 a render you can afford to run the rough version and the layered version side by side and keep whichever lands.

Where to go next

Try a prompt in the playground →
Writing Prompts for Explicit AI Images — Nureta Guides