Consistent Character Across Shots

Build a reusable identity packet, choose the right image mode, and review continuity at each cut rather than trusting a repeated name.

Continuity starts before the video prompt

A character can remain recognizable in an opening portrait and still drift when the camera turns. Plan the evidence each shot needs: the same face, hair silhouette, clothes, accessories and body proportions. This article focuses on continuity across separately generated shots. For a general upload walkthrough, see the Reference mode guide.

According to ByteDance’s Seedance overview, images can guide generated video. That is background for using visual evidence, not a promise of exact identity. Your job is to remove conflicting evidence and make the same approved design available whenever a new shot begins. A repeated character name is an organizational label, not stored model memory.

Prepare a clean identity packet

Choose an approved face image with even light and visible features. Add a compatible side or three-quarter view when the next scene needs that angle. Include a full-body view if clothing, shoes or body proportions carry the story. Prefer a neutral setting that does not compete with the location you intend to generate.

A character sheet should show the same person in the same outfit throughout. Remove contradictory accessories, alternate hairstyles and tiny decorative panels. Keep labels outside the visual design when possible, so a panel heading does not become part of the generated scene. An image containing several views is still an image file, not several independent upload slots.

Write an identity paragraph beside the files: “Mara has a rounded face, dark curly bob, olive work jacket, cream shirt and plain canvas shoulder bag. Preserve her face, hairstyle and clothing.” Give it a stable filename and reuse the exact wording. Keep emotional state and lighting outside that paragraph; those can change between shots without redefining the person.

Reference mode and first frames solve different problems

Seedance 2.5 and the Seedance 2.0 variants accept 1–9 reference images on SeaDance. MiniMax H3 and H3 Max have the same Reference image-count range. Use this route when images describe identity, wardrobe, props or a setting without needing to reproduce their exact opening composition. Describe each image’s intended role in plain language.

In Image to Video, the first uploaded image defines the opening frame. Seedance and the H3 family also accept an optional last image in this mode. Kling 3.0 accepts a single first image and has no separate Reference mode on SeaDance. A finished opening frame is suitable when face placement and screen direction must already be settled.

Kling 4.0 is available in the model directory; check the live quote in the generator. According to Kuaishou, its reference approach includes subject images and character sheets. Follow the image inputs shown by SeaDance; the manufacturer’s broader keyframe controls are not available on SeaDance. Do not assume another model’s upload limit applies.

Budget for continuity tests

Continuity testDuration / resolutionGeneration quote
Seedance 2.5, one reference image5 seconds / 480p11 credits / $2.21
MiniMax H3, one reference image5 seconds / 768P9 credits / $1.81
Kling 3.0, first frame5 seconds / 720p13 credits / $2.61

All dollar equivalents allocate prepaid credits at the Basic monthly rate: $29.90 for 149 credits, or approximately $0.2007 per credit. They are not separate checkout prices. See pricing for plans and packs. New accounts start with 0 credits. Each new attempt is charged again, including a completed take you dislike; a failed generation returns its generation credits. That is different from a refund of a subscription payment. See the refund policy.

Reuse identity; change only the shot instructions

Open every shot brief with the same identity paragraph, then describe the new setting, action and framing. For example: “Mara, as defined in the character image, stands beside the workshop door. Medium shot from inside the room. She pulls the door shut and looks toward the table. End with her shoulder bag still on her left shoulder.”

For the following shot, retain the identity text and make the new composition explicit: “Mara, with the same face, hair, jacket and shoulder bag, stands at the workshop table. Close view of her face and upper torso. She lowers her gaze to the folded note.” Do not silently rename the bag or change its color while asking for consistency.

Avoid instructions such as “continue from before” without restating the visible state. A separate generation needs its own context. Record which hand holds the prop, which side the person faces, and whether a door is open. These details are easier to review than a broad request to keep everything cinematic.

Inspect drift where the cut will happen

Compare the last usable pose of the outgoing shot with the opening of the next. Look at face shape before color grading, and inspect hairline, collar, sleeves and accessories. A lighting change may be intentional; a new jaw shape or missing bag is a continuity error for this brief. Review moving frames as well as the first still.

If identity changes after a turn, provide a compatible angle and reduce the turn’s complexity. If wardrobe changes, remove conflicting outfit images and keep the same clothing phrase. If background details leak into every shot, replace a busy portrait with a cleaner identity reference. Change the evidence responsible for the mismatch before lengthening the negative instructions.

For a first-frame route, repair the opening still before animating it. A malformed hand or incorrect jacket already present in the image is a poor foundation for motion. You can prepare stills separately with the image generator, but those are independently charged outputs. Existing approved photographs can also be used when you have permission.

Keep a continuity log with the approved files

For each scene, save the reference filenames, repeated identity paragraph, prop state, screen direction and the reason a take was accepted. When replacing a file, note which later shots need another review. This keeps a prettier new portrait from accidentally changing the character halfway through the sequence.

Use a small action test before committing to the entire scene: a look toward the camera, a turn toward a door, or a hand reaching for a familiar object. Approve the face and outfit in motion first. There is no guaranteed identity lock, and increasing the number of references cannot resolve references that disagree.

Frequently asked questions

Does using the same name preserve a face?

No. A name helps organize your prompt, but each generation needs visual and textual evidence. Reuse approved images and the same description, then inspect the result in motion.

Should every reference show the whole scene?

No. Clean identity images suit face and wardrobe guidance. A finished full scene suits an opening-frame route when the composition itself matters.

Can I upload nine images to Kling 3.0?

No. Its SeaDance image route accepts a single first frame. The 1–9 image Reference range belongs to the Seedance and H3 families described above.

Are more reference images always better?

No. Extra images can introduce conflicting designs. Choose the smallest compatible set that explains the person and the angles your scene actually needs.

Sources

Official background checked November 5, 2026. SeaDance settings and billing in this article follow its current product configuration, rather than the full range of manufacturer features.

Related articles