Building an NSFW Character Sheet with AI

A character sheet is a small, canonical set of images of one person — a clean front view, a profile, a full-body shot, maybe a few key poses — that you reuse as reference for everything else you make. Build it once and every later image, and every later video, can stay on-model instead of drifting into a new face each render.

The workflow has three moves: generate a base character you like, lock that identity with reference images, then expand it into the poses and angles you need. This guide walks each step on the seahorse-image model.

Step 1 — Generate a base character#

Start text-only. Write a prompt that fully describes the person — body type, adult age descriptors, face, hair, skin, and any distinguishing marks like tattoos or freckles you want them to keep. Render several variations in the playground and pick the one that best matches the character in your head. This first render defines who they are, so spend a few iterations here; it's the cheapest place to change your mind at $0.05 a render.

Step 2 — Lock identity with references#

Once you have a base you like, turn it into references. Render a small canonical set — front, three-quarter, and full-body — of that same person, keeping the prompt's identity details fixed. These become your reference images. Pass them back into future tasks as image_url items, and the model carries the face, body, and tattoos forward. The mechanics of how identity and markings carry are covered in using reference images for consistent faces and bodies.

If your references aren't already hosted, upload them once via the presign flow and keep the returned assetUrls handy — you'll reuse the same URLs across every image in the set.

Step 3 — Produce poses and angles#

Now hold the references constant and vary only the prompt. Each new pose, angle, or act is a request with the same reference set and a different text line. This is what fills out the sheet — standing, kneeling, close-up, from behind, explicit or clothed — all recognizably one character.

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": "the same woman kneeling on a bed, arched back, three-quarter view from behind, warm lamplight" },
      { "type": "image_url", "image_url": {"url": "https://.../ref-front.jpg"} },
      { "type": "image_url", "image_url": {"url": "https://.../ref-profile.jpg"} },
      { "type": "image_url", "image_url": {"url": "https://.../ref-body.jpg"} }
    ],
    "size": "2K"
  }'

Send up to nine references. Two or three sharp ones are usually enough; add more angles if a specific feature keeps drifting.

Step 4 — Reuse the sheet everywhere#

A finished character sheet isn't just for stills. The same reference images that keep an image set consistent also drive video — feed them into image-to-video to put your locked character into motion without re-rolling their face. One sheet becomes the identity backbone for an entire body of work across images and clips.

Keeping the sheet clean#

• Use sharp, well-lit references. A blurry or heavily shadowed reference gives the model less to hold onto.
• Keep faces front-facing where you can — a clear front view anchors identity better than an obscured one.
• Re-canonicalize occasionally: if a later render nails the character even better, promote it into your reference set.
• Store your reference URLs somewhere reusable so every request pulls from the same locked set.

Where to go next#

• Deeper on the mechanics: using reference images for consistent faces and bodies.
• Sharper prompts for each pose: writing prompts for explicit AI images.
• Put the character in motion: image-to-video.