A character can look right in the first frame and still change as the video unfolds. Their face may shift during movement, their hairstyle may change after a cut, or a second copy of the same character may appear in the scene.
For teams making a series of ads or recurring videos, character consistency needs more than a line saying “keep the same person.” It helps to give each reference a clear job, keep the scene manageable, and review the whole clip before using it.
Start with references that are easy to read
If one image combines a portrait, a full-body pose, outfit details, and several views of the character, the face can become a small part of a busy reference. BytePlus’s Seedance 2.0 prompt guide identifies this as a possible cause of inaccurate face matching.
A more focused setup is:
- Image 1 — facial identity: a clear, close view of the character’s face.
- Image 2 — appearance and styling: a single-person image showing the intended hairstyle, clothing, and overall look.
- Image 3 — another essential element: a product or setting reference, if the scene needs one.
Tell Seedance what to take from each image. Avoid giving several images overlapping roles without explaining which details matter most. For character references, BytePlus advises against multi-view character sheets when they could be interpreted as several subjects.
These steps can improve the clarity of a request. They cannot guarantee that every generated frame will match its reference.
Define the character once, then keep the name consistent
Give the character a short, stable description and use the same name throughout the prompt. Separate what should stay fixed from what can change between scenes.
| Keep consistent | Change when the story calls for it |
|---|---|
| Facial identity and distinctive features | Location and background |
| Hairstyle and signature accessories | Action and expression |
| Clothing, if the campaign requires it | Framing and camera direction |
| Overall visual style | Dialogue and scene purpose |
For example, a presenter might appear in a bathroom, a studio, and a café while retaining the same recognizable face and hairstyle. Clothing can change deliberately; an unexplained change halfway through a shot is something to catch during review.
Prompt example: introduce a fictional presenter
The example below uses Maya, a fictional digital character. The image labels describe the intended reference order; adapt them to the interface or API used for your request.
Prepare:
- Image 1: a clear facial reference of Maya.
- Image 2: a single-person styling reference showing Maya’s cream blouse and hairstyle.
- Image 3: the unbranded product bottle.
Example prompt:
Create a short skincare introduction featuring one presenter, Maya. Use Image 1 for Maya’s facial identity, Image 2 for her cream blouse and hairstyle, and Image 3 for the product bottle. Maya is the only presenter in the scene.
Shot 1: Medium shot at a bathroom counter in soft morning light. Maya stands beside the bottle, looks toward the camera, and gives a small smile. Keep the camera still.
Shot 2: Cut to a closer view. Maya slowly lifts the bottle to chest height, keeping her face and the bottle visible. She says in English, {This is where my morning begins.}
Shot 3: Return to a medium shot. Maya places the bottle on the counter and looks toward the window.
Keep Maya’s facial features, hairstyle, and blouse consistent across the shots. Show only one Maya in each frame. Use quiet room ambience without background music. Do not add captions or extra brand graphics.
The prompt assigns each image a role and describes the shots in order. It also gives the reviewer specific details to check: Maya’s face, her clothing, the bottle, and the transition back to a medium shot.
Reuse the character without repeating the whole scene
For the next video, keep Maya’s accepted references and change the setting and action. A shorter prompt is often easier to assess than a new brief that changes every detail at once.
Create a new scene featuring Maya. Use Image 1 for her facial identity and Image 2 for her cream blouse and hairstyle.
Maya sits at a small café table beside a window. Begin with a medium shot as she looks down at a closed notebook. Slowly move the camera closer as she looks up and smiles at someone outside the frame.
Keep Maya’s appearance consistent with the references. She remains seated, and no other person is clearly visible. Include quiet café ambience and no dialogue.
Compare the café clip with the bathroom clip side by side. Does Maya still read as the same character when the lighting, background, and action change? If the answer is unclear, record where the difference appears before changing the prompt or references.
Review and reuse the result
Watch beyond the opening frame. Check the face during movement, both sides of each cut, and the final frame. If the character handles an object or turns away from the camera, inspect those moments too.
Keep the model version, generation settings, reference files in order, exact prompt, and accepted output together. When the next result changes unexpectedly, this record makes it easier to see whether the difference came from the references, the prompt, or the scene itself.
A useful review note is specific: “Maya’s face changes after the close-up cut” gives the team a starting point. “The character looks wrong” does not.
Using real people as character references
The fictional example above should not be read as an instruction to upload an ordinary portrait of a real person directly. Seedance applies restrictions to face-containing input assets. Real-person use requires an eligible, consent-verified asset workflow and the applicable permissions.
LSF can help teams plan reusable digital-character references and, where available, the private real-human asset workflow. If you are building a recurring presenter or video series, bring a representative scene and the references you intend to use. We can help structure a test around the continuity your project needs.
Explore Seedance 2.5 through LSF or request a test account.
For more on organizing visual and audio references, read our Seedance 2.5 multimodal prompting guide.
References: BytePlus Dreamina Seedance 2.0 series prompt guide and BytePlus guide to creating portrait videos with Seedance. The examples and review process in this article are LSF’s practical suggestions, not tested performance guarantees.