When an AI short series or a branded character video falls apart, it is rarely the motion that breaks the spell. It is the moment your lead walks into the next shot wearing someone else's face. Small things count too: a hairstyle that grew an inch, an apron that switched colors, a hair clip that quietly vanished between cuts. Any of them is enough to make a sequence feel like a different production.
Seedance 2.0 takes image, video, audio, and text as references and can generate multi-shot content in one pass. But keeping a character stable is not something you can outsource entirely to the model. It takes fixed reference assets, prompts that repeat themselves on purpose, and a checking habit that runs after every single shot.
Here is how that works in practice, built around a fictional barista and a three-shot spot. One honest caveat first: this is a reference-image and prompting guide. The example frames are illustrations, not measured Seedance 2.0 outputs.
Why Faces Drift in the First Place
Character consistency means the same person stays recognizable across shots, actions, and scenes. Identity drift is what happens when it does not: facial proportions shift, accessories disappear, hair gets longer, clothing changes color. A single clip can look perfectly fine on its own and still read like recasting when you line it up with the others.
ByteDance launched Seedance 2.0 in February 2026 with multimodal references, multi-shot generation, and improved instruction following and consistency control. Worth reading carefully, though: improved control is not a guarantee. If each shot gets a different description of the person, or references that quietly contradict each other, the output will drift no matter how capable the model is.
Build the Character Sheet Before Anything Moves
Decide what cannot change before you write a single prompt. For this exercise, the character is a barista named Nora:
Face: oval shape, natural brows, soft features.
Hair: dark brown, chin-length bob, side-swept fringe, one silver hair clip.
Outfit: cream-white shirt under a dark green apron.
Detail: round copper buttons on the apron straps.
Look: photographic realism, soft natural light.
That list is the contract. Every shot inherits from it, and every review checks against it.
Two references beat a pile of references
More inputs are not automatically better. ByteDance's own prompt guidance for the series suggests starting with two separate images: a face close-up that locks identity, with a neutral expression and a clean background, and a full styling shot that carries the hair, outfit, accessories, and proportions. Two files, two jobs.
What you should not do is stitch a front view, a side view, and a back view into one collage and upload that. The official warning is blunt about why: a multi-angle sheet can read to the model as several different people in one frame, which raises the odds of a face swap or a duplicate character wandering into your shot.
Three Shots, One Person
The spot is simple: a face, a gesture, a turn. Each shot has one job, and the prompts repeat the fixed traits verbatim instead of paraphrasing them.
Shot one: front close-up, establish the face
Nora faces the camera with a light smile while the lens pushes in slowly. The prompt names both references (image one for the face, image two for the styling), then lists the protected traits: features, short hair, silver clip, outfit unchanged. Photographic style, soft daylight, exactly one person in frame. Because the shot exists to lock the face, big head turns and fast motion stay out of it.
Shot two: counter medium, hold the wardrobe
Same two references, new setting. Nora stands behind a wooden counter and lifts a white coffee cup, eyes following it. The prompt keeps repeating the same anchor list, face, bob, clip, cream shirt, green apron, and adds what must not happen: fixed camera, natural motion, no new people, no costume changes. Review focus here: did the apron color, the straps, and the hair carry over from shot one?
Shot three: by the window, test the angle
The hardest one. Nora stands at a window, body half-turned, then looks back over her shoulder as afternoon light comes through the glass. Same references, same protected list, plus an explicit instruction not to change the person's identity. This is where profile proportions, hair length, and accessory placement get checked, because an angle change is where drift likes to hide.
Then the step most people skip: put the three clips together and watch them as a sequence. Drift that survives a solo review almost never survives the lineup.
Fixing a Drifted Face, Outfit, or Hair
When something slides, resist the urge to rewrite the whole prompt. Identify which trait broke, then adjust the reference or the instruction that governs it:
Face changed by shot three: restate the priority of the standalone face close-up and remove competing images from the input.
Green apron turned black: keep reusing the exact same styling photo and name the color explicitly every time.
Bob suddenly grows: add length and style limits to the prompt, and shrink the head-turn range.
Hair clip vanishes or migrates: use a styling shot where the accessory is clearly visible, and check for it shot by shot.
A second Nora appears: almost always the collage problem. Go back to single-person references and state the character count outright.
The difference between a vague instruction and a workable one is mostly specificity. A line like "Nora walks to the window and looks back, keep her pretty" tells the model nothing about what to preserve. "Use the same face and styling references. Nora turns her head slowly. Keep the original face shape, feature proportions, chin-length bob, silver clip, and dark green apron. No extra people, no costume changes. Fixed camera, small motion" gives every trait a guard rail.
If the face is the thing that broke, look at the face reference first: is it actually clear, is the background competing with it? Minor accessory differences can sometimes be patched in editing. A visibly different face is usually a regenerate, not a retouch. All of this lowers the odds of drift. None of it guarantees two identical shots.
Before You Export
Check every clip for face, hair, outfit, accessories, and color tone before anything ships. One shot with an obvious face swap ruins a sequence no matter how polished the other two look, and if you are assembling inside a workflow tool, confirm which model version your entry point actually runs. Reference limits, clip length, resolution, and credit costs differ between versions, so specs from one interface do not transfer cleanly to another.
Questions That Come Up
Can a locked character change clothes?
Sure, as long as the change is written into the story. State which shot the costume switches in, keep the face and hair settings intact, and manage references for each look separately. What you want to avoid is one reference set quietly arguing with another.
Does reusing the same seed prevent face swaps?
No. A seed is one input among many, not a continuity lock. You still need the same character references, the same model settings, and the same prompt anchors, and not every interface even exposes seed control. Check yours before building a workflow around it.
Why did the reference generate two identical characters?
The usual culprit is a multi-angle collage the model read as multiple people. Switch to separate single-person references and say plainly, in the prompt, that the frame contains one character.
The Order of Operations
Consistency work belongs at the reference stage, not in post. Lock the face and the styling first, then let each shot carry a different job: one establishes identity, one proves the wardrobe survives motion, one stress-tests the angle. When something breaks, change only the part that broke, so you always know which decision earned its keep. Start with a short clip and a small motion range, and only add camera ambition once the character survives the lineup.
