Characters

What a character is made of

• Name — display label. Used as a placeholder hint in prompts (so naming your penguin “Rasputin” doesn’t confuse the QA model into looking for a historical figure).
• Description — your seed text. A 1–2 sentence sketch is fine; SteadyShot auto-expands this into the rich descriptor with Claude.
• Extracted descriptor — the rich, prompt-ready identity (face, hair, body, clothing, distinctive features). This is what actually gets baked into every scene prompt. You can edit it directly or refine it with the AI co-pilot.
• Style notes — art register (e.g. “Pixar 3D”, “watercolor”). Stays constant across scenes unless overridden.
• Source image (optional) — the original photo or illustration. If present, it’s used as the visual anchor for the first reference and (in PuLID mode) as the face embedding for every scene.
• Reference sheet — generated views of the character (front, side, back, expressions). Multiple references give the image model more angles to triangulate identity from.

Draft vs locked

Every character starts as a draft. Drafts are fully editable: name, descriptor, references — all mutable. The tradeoff is that drafts can’t do anything else:

• Scene generation is blocked.
• Library publishing is blocked.
• Marketplace listing is blocked.
• External API calls (MCP, Figma, Photoshop, REST) 409 with a “lock first” error.

Locking freezes the character. From then on every generation anchors against the exact same descriptor + references, so consistency stays high across scenes. You can unlock if you need to edit — but you’ll start a new revision and previously generated scenes will reflect the pre-edit version.

How identity stays consistent

SteadyShot uses a layered approach to keep the same character recognizable across scenes:

1. Reference images as anchors. Up to 4 are passed to the image model per generation. The front view is always pinned; others are chosen based on keywords in the prompt (e.g. a prompt with “smiling” pulls in the “happy” reference).
2. Rich descriptor in the prompt. The extracted descriptor (not just your seed text) is appended to every scene prompt with explicit instructions like “match colors and clothing exactly”.
3. Per-call cache. The character’s identity descriptor + references are warm-cached server-side, so a fast sequence of scenes doesn’t pay the cold-start cost.
4. Optional scene chaining. Pass previous_image_id to thread the prior scene as the highest-priority anchor — cuts cross-scene drift in multi-page flows like storybooks.
5. Vision-based QA. Every generated scene is scored against the canonical references with Claude vision; if it drifts too far the system retries with feedback baked into the prompt.

Imported characters

Two ways to add a character without creating it from scratch:

• Free community library — browse /library and import any published character into your account. Zero credits.
• Paid marketplace — buy a creator’s character for credits. See Marketplace.

Both paths produce a fully-owned clone in your library (descriptor + reference sheet). The original creator retains resale rights — you can use the imported character for generations but can’t re-list it.

Doing it via the API

# Create
curl -X POST https://steadyshot.ai/api/characters \
  -H "Authorization: Bearer ss_live_xxx" \
  -F name="Luna" \
  -F description="A round orange tabby with green eyes" \
  -F style_notes="watercolor"
# → { "id": "...", "lock_status": "draft", ... }

# AI auto-fill the descriptor
curl -X POST https://steadyshot.ai/api/characters/{id}/auto-fill-descriptor \
  -H "Authorization: Bearer ss_live_xxx"

# Generate reference sheet (one POST per label)
curl -X POST https://steadyshot.ai/api/characters/{id}/references \
  -H "Authorization: Bearer ss_live_xxx" \
  -H "Content-Type: application/json" \
  -d '{"label":"front","has_source":true}'

# Lock
curl -X POST https://steadyshot.ai/api/characters/{id}/lock \
  -H "Authorization: Bearer ss_live_xxx"

Full schema at REST API.

Examples

Three concrete recipes you can copy as a starting point. Each one has a different intent — children’s storybook, brand mascot, and photo-real OC — and uses a different descriptor + art style profile.