Seedance 2 Consistency: Characters and Products

Seedance 2.0 consistency starts with a controlled reference packet, not a repeated adjective. Give one image responsibility for character identity, another for exact product geometry when needed, and separate sources for environment or motion. Then lock the facts that cannot change, reduce unseen camera angles, and log one repair per retry.

ByteDance says the upstream model can accept as many as nine images, three videos, and three audio clips. Oakgen's current Seedance 2.0 text-to-video form exposes the same reference categories; that does not mean you should fill every slot.

The four-file packet

Start with fewer sources than the maximum:

Add an audio reference only when tempo or sound character changes the edit. Oakgen's current Seedance form accepts reference audio but exposes no audio-generation toggle; inspect the returned file and plan the delivery soundtrack separately.

Character reference checklist

• Face and hair edges are sharp.
• Wardrobe is fully readable and does not merge into the background.
• Hands involved in the action are visible or the action starts after contact.
• The pose leaves room for the requested move.
• The reference does not contain another prominent face.
• A second view adds genuinely missing profile or full-body evidence.

Product reference checklist

• Silhouette, cap, hinge, handle, controls, and material are readable.
• Angles cover the sides the camera will reveal.
• Approved logo and label files remain separate for post-production.
• Color is documented alongside the lighting condition.
• Scale has a believable anchor.
• No crop removes a critical product edge.

The product video prompt guide covers background, crop, source, and delivery-ratio issues.

Write one continuity ledger

For a creator holding a travel mug:

CHARACTER LOCK: same woman; oval face, short black bob, small gold stud earrings,
gray wool coat, red scarf, natural skin texture. No identity, age, hair, jewelry,
or wardrobe change.

PRODUCT LOCK: same navy travel mug; tapered body, silver rim, black lid, left-side
handle, matte finish, fixed proportions. No duplicate mug, moving handle, new control,
redrawn label, color change, or scale change.

WORLD LOCK: quiet morning train; cool window light; warm carriage reflection;
same seat, table, weather, and time.

Keep these lines unchanged between retries. If you rewrite the identity facts while testing camera movement, you have changed two variables.

Assign reference roles in the prompt

Reference image 1 controls only the woman's identity and wardrobe.
Reference image 2 controls only travel-mug geometry, material, and color.
Reference image 3 controls train layout and lighting.
Reference video 1 controls only camera height and slow forward pace.

ByteDance says Seedance can reference composition, motion, camera movement, visual effects, and sound from input assets. Role assignment tells it which part matters to this shot.

Run a stability test before the full scene

[PASTE CONTINUITY LEDGER AND REFERENCE ROLES]

Five-second medium shot. The woman sits beside the train window holding the mug
with both hands already in place. She takes one natural breath and looks from the
window toward the camera. Slow straight camera push forward. No speaking, drinking,
lid movement, hand repositioning, head turn beyond 20 degrees, or new objects.

Review face, hair, earrings, coat, scarf, mug silhouette, lid, handle, material, and scale. Only expand the action after those fields hold well enough for the delivery.

Expand one difficulty at a time

This table is a protocol, not a claim that every pass will succeed. Record what happens.

Consistency scorecard

Score each field from 0 to 2 at the first frame, midpoint, and final frame.

Keep the score with the generation settings and prompt. Do not turn one clip into a universal benchmark.

Repair sequence

1. Find the first drifting frame.
2. Name the newly introduced demand at that frame.
3. Remove or reduce that demand.
4. Add missing angle evidence only if the shot needs it.
5. Keep all other inputs fixed.
6. Compare the same fields in the scorecard.

For the broader workflow, continue with the Seedance prompting guide.

Seedance-specific control boundaries in Oakgen

Oakgen currently exposes two separate Seedance paths:

• text-to-video with optional image, video, and audio references;
• image-to-video with one required opening image and an optional last image.

Use the text-to-video path for an assigned reference packet. Use image-to-video when the opening composition is the dominant source of truth or the ending needs a last frame. The control-mode decision guide compares both.

This article covers Seedance 2.0. The existing Seedance 2.5 character-consistency guide covers the newer 2.5 workflow and should not be treated as proof of identical controls in 2.0.

Common mistakes

Adding more identity images after every failure. Remove conflicting references first. A larger inconsistent set weakens ownership.

Testing character, product, camera, dialogue, and lighting together. Pass a short stability shot before adding interaction.

Generating exact labels. Preserve a tracked surface and composite approved artwork later.

Using a motion reference with cuts. One short move is easier to assign than a montage.

Calling a visually similar person “consistent.” Decide recognition fields before generation and score those fields.

FAQ

How many images can Seedance 2.0 use?

ByteDance's official launch states as many as nine image references, plus three video and three audio clips. Oakgen's current text-to-video form exposes matching image, video, and audio reference fields with host-specific limits.

Should I use all nine images for one character?

No. Use only views that add missing identity evidence. Conflicting lighting, age, hair, wardrobe, or face detail can weaken the result.

Can I keep both a person and product consistent?

Yes, but it is a harder shot. Give character and product separate sources and locks, then start with contact already established.

Why does the product change during a sip or pickup?

Hand contact adds geometry, force, occlusion, and scale problems. Simplify the grip, widen the frame, or split the contact into its own shot.

Can I use a last frame for consistency?

It can anchor the final composition in Seedance image-to-video, but both endpoints must show the same subject and a believable transition.

Does Seedance 2.0 produce audio in Oakgen?

Oakgen currently accepts audio references in text-to-video but exposes no audio-generation toggle. Provider requests can use endpoint defaults, so verify sound in the returned file.

Continue the workflow

• Read the Seedance 2 complete guide.
• Build a 15-second multishot sequence.
• Diagnose other failures with the Seedance prompting guide.
• Compare Seedance, Kling, and Veo by creator job.

Sources

• ByteDance Seedance 2.0 official launch
• ByteDance Seedance 2.0 model overview