Seedance 2 Prompts for Consistent Characters: A Production Prompting Guide

By Character App Team · · 12 min read

Most Seedance 2 prompt guides hand you generic formulas that fall apart the moment you need the same face twice. This is the character slice: how we prompt Seedance 2 in production for consistent characters, why the reference does most of the work, and copyable patterns for the drama scenes you actually shoot.

How Seedance 2 reads a prompt

We run Seedance 2 in production at Character App, and the single most useful thing to understand before you write another prompt is how the consistency workflow actually works. For consistent characters, you run Seedance 2 image-to-video: you hand it a starting frame rendered from your locked Cast, and it animates that frame. When you work this way, your prompt describes the motion, the camera, and the mood, while the starting Image supplies the face, the wardrobe, the lighting, and the art direction.

That distinction reorganizes everything. If the appearance already lives in the frame, then a prompt crammed with facial detail is not helping — it is competing with the picture the model was handed. The prompts that stay consistent are the ones that trust the image to carry identity and spend their words on what the image cannot show: what the character does over the next few seconds, and how the camera watches it happen.

Practically, Seedance 2 rewards restraint. It renders cinematic output with native audio and gives you clips from about 4 to 15 seconds (as of July 2026), which is exactly enough room for one clean beat. A clip is not a scene edit — it is a single moment. Write it like one.

The reference-first principle

Consistency is not a prompting trick. It is a decision you make before you write the prompt: lock the identity in a reference, then stop describing it.

A Cast in Character App is a reusable saved reference — built from a single photo or a text prompt — that gets fed into every render as the identity authority. When you assign a Cast to a scene, the still Image the model animates already looks like your character. From there, the fastest way to cause drift is to re-describe the person in words. "A woman with sharp cheekbones, auburn hair, green eyes, a leather jacket" gives the model a second opinion to argue with, and the argument shows up on screen as a face that shifts shot to shot.

So the reference-first rule is simple: name the role, not the appearance. Write "she" or "the detective," describe what she does, and let the locked reference own how she looks. The one exception is anchoring to the frame — if the starting Image shows a red coat, you can say "keep the red coat" to hold it. What you never do is invent new features the reference does not have.

Start from the right frame

Because Seedance 2 animates a starting Image, that image is your most powerful consistency control — more powerful than any prompt. The frame owns appearance, composition, lighting, and art direction, so a weak starting Image cannot be rescued by clever wording. Get the frame right and the prompt gets easy.

When you render the starting Image from your Cast, aim for a frame that is close to the moment you want to animate. If the beat is a slow turn to camera, start from a three-quarter angle so the turn has somewhere to go. If it is an emotional close-up, start tight on the face. Starting far from the target pose forces the model to invent a large motion, and large invented motion is where faces slip.

Prompt anatomy for drama scenes

Our production prompts follow a six-part order that maps to how the model resolves a shot: Subject, Action, Environment, Camera, Style and Lighting, Constraints. You do not need every part every time, but the order keeps the important information early and the guardrails at the end.

The two parts that make or break a character clip are Action and Constraints. For Action, the reliable pattern is one readable physical movement followed by one small settled reaction — a turn and then a held look, a step forward and then a breath. Stacking three actions into a four-second clip is how you get morphing and extra limbs. For Constraints, you tell the model what not to do: no cuts, no camera shake, no morphing, and hold the face and wardrobe exactly as framed.

Here is the skeleton we fill in, delivered as copyable text you can paste into a Story scene and adapt:

[Subject: role or pronoun only] + [Action: one physical action, then one small reaction] + [Environment: place and time of day] + [Camera: framing, and either locked or one slow move] + [Style/Lighting: light direction and depth of field] + [Constraints: no cuts, no morphing, preserve the face and wardrobe as framed].

Notice there is no facial description anywhere in the skeleton. That absence is the point — the Cast reference fills it in every time, which is what keeps the character the same across a whole run of scenes.

Words to cut, words to keep

After enough production runs you develop a feel for which words earn their place in a character prompt and which words invite trouble. Two quick lists to keep next to your keyboard:

Copyable patterns for the scenes you actually shoot

These four cover most of what a vertical drama needs: a confrontation, a turn, a burst of motion, and an emotional beat. Each one is written for an image-to-video render, so identity is assumed to be carried by the assigned Cast. Copy them, swap the specifics, and keep the constraint line intact.

The dialogue beat

A single character delivering a line to someone off-frame. You are directing the performance, not scripting synced speech — write the emotional beat and let the acting read.

Prompt: "She leans in slightly and delivers a line with a tight, controlled expression, then presses her lips together and holds the other person's gaze. Small kitchen at night, one warm practical lamp off to the side. Camera holds a locked medium close-up at eye level. Soft key from the left, shallow depth of field. No cuts, no camera shake, keep her face, hair, and jacket exactly as framed."

If you need exact spoken words, treat that as an edit-stage decision — these models animate the delivery, they do not reliably lip-sync a script. Native audio gives you room, breath, and ambience under the beat; precise dialogue you cut in your own edit.

The reveal

The workhorse of episodic hooks — a recognition landing on a face. One turn, one shift of expression, one slow push.

Prompt: "He turns his head slowly toward the doorway and his expression shifts from calm to alarm as recognition lands. Dim hallway, single overhead light. Camera pushes in slowly from a medium shot toward a tight close-up. Cool low-key lighting, hard shadow across half the face. No cuts, no morphing, preserve his face and coat throughout."

The chase

Motion is where consistency is most fragile, so keep the action singular and the background quiet. One clear movement, a tracking camera, and explicit limits on everything else.

Prompt: "She sprints toward the camera, arms pumping, glancing back once over her shoulder. Narrow neon-lit alley at night, wet pavement. Camera tracks backward, staying with her at chest height. High-contrast neon key, background action kept minimal. No extra people, no cuts, no limb distortion, keep her jacket and hair as framed."

The emotional close-up

The quietest and most consistency-friendly shot in the format. Almost no movement means almost nothing for the model to break — which is why these tend to render cleanest.

Prompt: "A single tear forms and slides down her cheek as she blinks slowly and her jaw tightens. Quiet bedroom, soft window light. Camera holds a still, locked extreme close-up. Gentle side light, shallow focus on the eyes. No camera movement, no cuts, hold the exact face and makeup from the frame."

Common failure modes and fixes

When a Seedance 2 clip comes back wrong, it is almost always one of a handful of causes — and each has a direct fix that lives in the prompt, not in the model.

What a locked Cast handles so your prompt does not have to

The reason these prompts can be so light on description is that a locked Cast is doing the heavy lifting upstream. Because you run Seedance 2 image-to-video, and the starting Image it animates is rendered from your Cast reference, the face, the wardrobe, and the visual style arrive already correct. Your prompt inherits a solved problem.

That is also what makes a series possible. Assign the same Cast across every scene of a Story, and each shot animates from a matching identity — so a character you plant in episode one is still recognizable in episode twenty. A Universe can pin a shared Look on top of that, so the whole run holds one visual direction without you re-specifying it per prompt. You can go deeper on the mechanics in our model-agnostic consistency breakdown at /consistent-ai-characters, or start building a character at /ai-character-creator.

The takeaway: consistency is architecture, not adjectives. Spend your effort building a clean Cast reference once, and every prompt after that gets shorter and more reliable.

The draft-to-hero workflow

Seedance 2 is a premium render, so you do not want to discover a blocking problem on a premium clip. The production move is per-scene model selection: draft cheap, then upgrade the shots that earn it.

We draft every scene of a Story with LTX Fast — the cheapest draft model — to test pacing, staging, and whether the action beat reads at all. LTX Fast is not the look you ship; it is the look you check. Once the blocking is right, we upgrade the hero scenes to Seedance 2 (or Seedance 2 Fast for a cheaper cinematic pass) and let the weaker connective shots stay on a lighter model. Because the prompt and the starting Image travel with the scene, your locked blocking survives the upgrade — you are swapping the engine, not rewriting the shot.

Every paid render shows its Credits price before you confirm, so you always know what a hero scene costs before you spend it. Plans start at $19.99/mo, and a free account covers daily Images so you can build and test a Cast before rendering any video (video generation is not part of the free tier). See the full roster at /pricing.

If you are deciding which engine deserves your hero Credits, we put Seedance 2 and Kling head to head on exactly this question — character consistency — in our production comparison at /blog/seedance-2-vs-kling-3-character-consistency.

Frequently asked questions

Do I need to describe my character in every Seedance 2 prompt? No — and you should not. The consistency workflow runs Seedance 2 image-to-video, so the starting Image rendered from your Cast already owns the appearance. Describe the action and camera; refer to the character by role. Re-describing the face is the fastest way to cause drift.

How long can a Seedance 2 clip be? Roughly 4 to 15 seconds per clip (as of July 2026). That is enough for one clean beat — one action and one reaction — which is exactly what a consistent character clip should contain.

Will Seedance 2 lip-sync my dialogue word for word? Treat it as a performance tool, not a lip-sync tool. Direct the emotional beat of a line and cut exact spoken audio in your own edit. Native audio adds ambience and presence under the shot.

My character still drifts even with a reference — why? Almost always a prompt problem: you contradicted the frame, packed in too many actions, or re-described the face. Match the starting Image, keep to one action, and let the reference carry identity.

Can I use these patterns for vertical drama? Yes. Seedance 2 renders 9:16 natively, so these prompts drop straight into a phone-first series. Build the series side of it with /vertical-drama-maker, and see the app-side prompt reference at /seedance-2-prompts.

What does a Seedance 2 scene cost? Every paid action shows its Credits price at confirmation, so there are no surprises. Plans start at $19.99/mo with a free tier for daily Images; the details are at /pricing.