Seedance 2.0 Character Consistency: A Full Workflow Deep-Dive.

Why Character Consistency in Seedance 2.0 Is Harder Than It Looks

Seedance 2.0 promises significantly improved character consistency across frames and shots. That promise is real, but it comes with a condition most users only discover mid-project: the model's content filter and its face-generation pipeline are not entirely separate systems. When the filter re-evaluates a generation involving a human face, it can interrupt the identity signal, producing a character who looks plausible but not the same person. The result is what field tests describe as "small, creeping changes that make a character feel like a different person across frames or shots." You get a character. You do not always get your character.

This is distinct from the filter avoidance problem covered in the companion article. Passing the filter is the door. Everything described in this guide is the room beyond it. Once your prompt clears content review, you still need a disciplined system to keep the model producing the same identity across every shot in a sequence. Without that system, you are relying on probability.

The problem surfaces differently depending on your discipline, but it surfaces everywhere. A game developer building cinematic cutscenes loses their NPC's silhouette between scenes: the character arrives in the medium shot with a jacket that has inexplicably changed cut, and their distinctive scar is gone by the third sequence. An automotive designer running a campaign sequence loses the driver figure's proportions between exterior and interior shots, making the two clips feel like different campaigns. A content creator building a branded persona for weekly social clips finds their character's hair colour drifts across the series, breaking the visual continuity that makes recurring content land.

In every case, the failure mode is the same: no system. A single clever prompt might produce one great shot. A system produces six shots that read as the same person. This guide builds that system, surface by surface, from the canonical character brief through to a repeatable Canvas workflow.

Building a Character Brief That Travels Across Shots

A canonical character brief is a fixed block of descriptive language that covers physique, clothing, distinctive features, and lighting relationship. It does not change between shots. It gets pasted verbatim into every generation prompt, every time. This is the single most important structural discipline in a multi-shot workflow, and it is the one most commonly abandoned.

How to Structure the Character Block

The optimal character block for Seedance 2.0 is ordered deliberately. Lead with what the filter accepts without friction: physical descriptors that are neutral and specific. Then anchor identity in clothing and silhouette before you describe the face. Silhouette is the model's most stable identity signal across shots. Face is its least stable. Layer in expression and pose last, as variable elements that can shift between scenes while the core identity holds.

The difference between a brief that survives shot changes and one that does not comes down to specificity. Compare these two descriptions:

Loose: "young woman in a red jacket"
Canonical: "woman, late twenties, straight black chin-length hair, oversized scarlet utility jacket with visible brass zips, matte skin, soft overcast key light from frame left"

The loose version gives the model creative latitude. Every new shot, it exercises that latitude differently. The canonical version removes almost all interpretive freedom on the identity elements. The jacket has a specific cut, a specific colour name, a specific hardware detail. The lighting relationship is fixed to a direction. The model can still make decisions about background, motion, and depth, but it cannot drift on the elements that define the character.

Seedance 2.0's architecture treats characters as persistent data objects rather than text-generated pixels when given sufficient reference signal. The canonical brief is how you build that reference signal in text form.

Storing the Brief in Write

Stensyl's Write surface is the right home for the canonical brief, and not just because it is convenient. Keeping the brief inside Write means it stays in the same workflow environment as the generation surfaces. You can open Write alongside Generate or Film in a split view, copy the character block directly, and paste without switching applications or hunting through a notes app. The brief also becomes a document rather than a snippet: you can version it, annotate which elements are locked and which are variable, and share it with collaborators inside the same project.

On Pro tier, Write gives you access to Claude Sonnet 4.6 and Claude Opus 4.7 for drafting and refining the brief itself. This matters because a capable writing model will catch vague language before it becomes a consistency failure. "Wearing a jacket" is vague. "Oversized scarlet utility jacket with visible brass zips at the chest and cuffs" is not. The model will surface that distinction if you ask it to.

The Most Common Error

The mistake that kills most multi-shot projects is adjusting the character brief between shots to "help" the model. The instinct makes sense: if shot two looks slightly wrong, you tweak the description to correct it. What you actually do is create a second character variant. Shot three is now generated from a different identity anchor than shot one. By shot five, you have a character family, not a character. Lock the brief before you generate anything. Never modify identity elements after the sequence has started.

Anchoring Visuals: Reference Images and Seed Discipline

Text alone is not enough. Seedance 2.0's consistency improvements depend substantially on visual reference: multi-angle collages showing the character from the front, profile, and three-quarter view under neutral lighting. The model uses these as identity data rather than interpreting the character fresh from text each time. A character who exists only in words is vulnerable. A character who exists in both words and a locked reference image is substantially more stable.

The First Clean Output Is Your Anchor

Do not use a mood board image as the visual anchor. Use the first approved generation. Run a short identity-check clip first: a four-second static or near-static generation using the canonical brief at its tightest. Evaluate it against every locked identity element. If the jacket hardware is visible, the hair length is correct, and the lighting relationship matches the description, that frame becomes the anchor image for every subsequent generation in the sequence.

This matters because a mood board image carries a different visual grammar than a Seedance output. The model will try to reconcile the two, and the reconciliation introduces variability. An anchor frame produced by the same model under the same prompt conditions is a cleaner input.

The Image-to-Video Flow in Generate

Stensyl's Generate surface handles the image-to-video flow directly. Upload the anchor frame, append the canonical prompt block to the generation prompt, and run the shot. On the next scene, upload the same anchor frame, not the previous shot's final frame, unless the character's state has definitively changed in a way you want to carry forward. The anchor frame is the constant. Only the motion instruction and scene context change between shots.

This approach is more stable than text-to-video for character work precisely because it removes the re-interpretation step. The model receives a visual identity rather than constructing one from language.

Seed Management

When Seedance 2.0 exposes seed values through the interface, lock the seed after the identity-check clip. The seed acts as the mathematical foundation of the noise pattern: reusing it across shot descriptions maintains the underlying geometric structure of the face. This is why two shots generated from the same seed with only the motion instruction changed look like the same person even when the text prompt is slightly imprecise.

When seed values are not exposed by the platform, the image-to-video anchor discipline described above is your primary consistency mechanism. There is no published workaround for unexposed seeds, and inventing one by guessing parameters will not reproduce the effect. Default to the visual anchor approach.

Storyboards Before Credits

Stensyl's Storyboards surface is the discipline that makes multi-shot consistency tractable before you spend a single credit. Board each scene before generating: define the required character states, camera positions, and expression changes across the whole sequence. Once you can see the full character arc on the board, you know exactly how many shots you need, where the identity-critical moments are, and where a costume or lighting change is intentional rather than drift. A motion designer building a character-driven loop, a film and set designer blocking a pre-vis sequence, and an exhibition designer animating a branded figure across multiple screens are all doing the same fundamental work here: making decisions before generation rather than discovering them expensively after.

Chaining Shots Through the Film Surface

Stensyl's Film surface is the practical home for multi-scene character work. It organises scenes in sequence and makes the relationship between shots visible, rather than leaving you managing a folder of disconnected files with names like "shot3_v4_final_FINAL.mp4". That visibility is not a cosmetic improvement. It is the structural difference between catching drift at scene two and catching it at scene eight.

Building the Chain

A concrete three-scene chain works like this. Scene 1 establishes the character: run the canonical prompt block with the anchor frame as the image input, full character in frame, establishing shot. Approve it at the clothing and silhouette level before touching Scene 2. Scene 2 cuts to a medium shot using the Scene 1 output frame as the image anchor, with a motion instruction that moves the camera rather than the character. Scene 3 introduces active motion using the same anchor discipline from the same source frame. The character block in the prompt does not change across any of these scenes. Only the camera and motion instructions change.

The inspection step is not optional. After each generation, compare the output against the anchor frame at the clothing and silhouette level. Check the jacket hardware. Check the hair cut. Check the key light direction. Catching drift at scene two costs one generation. Catching it at scene eight costs eight, and the repair requires returning to scene two anyway.

Credit and Concurrency Planning

Each generation draws credits from your active plan. Concurrent generation limits determine how many scenes you can run simultaneously. On Lite and Starter tiers, that limit is 1 concurrent generation. On Pro, it is 2. On Studio, it is 4. For a twelve-shot sequence on a Pro plan, this means you can run two scenes in parallel and should plan the generation order to take advantage of that without creating dependencies between simultaneously running scenes.

Count your scenes on the Storyboards surface before you begin generating in Film. A twelve-shot sequence that discovers a brief error at scene six and requires a full restart will cost significantly more credits than one that validated the brief in the identity-check stage. The planning step is credit management as much as it is creative discipline.

Film as an Audit Trail

The Film surface's scene structure doubles as a consistency audit trail. The sequence is visible in one place: you can scrub between scenes, compare character states at a glance, and identify where drift entered the sequence without opening multiple files. This is particularly valuable for collaborative projects, where a second reviewer needs to evaluate consistency without reconstructing the generation history from filenames.

Using Canvas to Systematise the Workflow

Film produces the sequence. Canvas makes the process repeatable. If you are running recurring character work, whether that is a weekly branded character series for a content and social team, or a game development studio iterating on an NPC across multiple asset batches, rebuilding the workflow from scratch each time is where consistency fails. Canvas builds the workflow once.

A Practical Canvas Layout

A character consistency Canvas connects three node types in sequence. The first is an LLM Chat node, which refines the canonical brief on demand. The second is an Image Generate node, which produces or updates the anchor frame. The third is a Video node, which runs the shot generation using the anchor frame and the refined brief as inputs. When you need to update the character's clothing for a new campaign phase, you adjust the brief in the LLM Chat node and the change propagates through to the generation step without you manually updating three separate prompt fields in three separate surfaces.

The LLM Chat node's model choice matters here. On Pro tier, Claude Sonnet 4.6 is the right default for brief refinement: it is precise with descriptive language and will flag ambiguous identity elements before they cause generation failures. On Starter, GPT-5.5 or Gemini Pro handle the same task competently. On Lite, GPT-5.4 mini and Gemini Flash are available and adequate for straightforward brief work, though they are less reliable at catching subtle vagueness in physical descriptions.

Tier	Monthly Cost	Concurrent Generations	LLM Chat Node Models Available
Lite	£10/mo	1	GPT-5.4 mini, Gemini Flash
Starter	£22/mo	1	GPT-5.5, Gemini Pro
Pro	£42/mo	2	Claude Sonnet 4.6, Claude Opus 4.7
Studio	£84/mo	4	Claude Sonnet 4.6, Claude Opus 4.7

The Limitation Canvas Cannot Fix

Canvas accelerates a working system. It does not fix a broken brief. If the canonical character block is vague, ambiguous, or internally inconsistent, running it through a Canvas workflow at scale produces consistent failures rather than consistent characters. The brief must be correct and validated against an approved anchor frame before you build the Canvas around it. Speed is the reward for correctness, not a substitute for it.

Diagnosing and Fixing Consistency Failures Mid-Project

Even with a disciplined system, failures occur. The three most common are silhouette drift, facial feature instability, and lighting relationship drift. Each has a specific cause and a specific fix. Knowing which failure you are looking at prevents you from applying the wrong correction.

Silhouette Drift

Silhouette drift is when the clothing shape changes between shots: the jacket becomes shorter, the collar changes structure, the fit shifts from oversized to tailored. This is a brief precision failure. The model has creative latitude on garment construction and is exercising it. The fix is adding more specific garment language to the canonical brief: not "oversized jacket" but "oversized utility jacket, dropped shoulders, hem below the hip, single breast pocket at chest left." The more construction detail the brief contains, the less latitude the model takes.

Facial Feature Instability

Facial feature instability occurs even when the prompt is passing the content filter cleanly. This is where switching from text-to-video to image-to-video makes the most immediate difference. The anchor frame carries the face as a visual datum rather than a verbal description. The model does not need to interpret "matte skin, straight black chin-length hair" because it can see them. If facial instability persists even on image-to-video, check that the anchor frame itself is clean: soft or motion-blurred frames produce ambiguous identity signal and should be replaced with a sharper still.

Lighting Relationship Drift

Lighting drift makes the same character read as two people. The face is structurally similar but the key light has moved from frame left to overhead, changing the shadow pattern on the face and making the character look different without any identity element actually changing. The fix is specifying the key light position in the canonical prompt rather than leaving it to the model. "Soft overcast key light from frame left" is a locked instruction. "Dramatic lighting" is an invitation for the model to interpret freely on every generation.

Using Ray to Diagnose Faster

When a failure mode is not immediately obvious, Ray is the right starting point. Describe the consistency problem in plain language: "my character's jacket silhouette is changing between scenes and I cannot work out whether the issue is in the prompt, the anchor frame, or the generation surface." Ray is built for exactly this kind of model and workflow decision. It will identify whether the fix lives in the prompt, the surface choice, or the reference image strategy, and direct you to the right correction without requiring you to test each variable independently.

Resist the temptation to regenerate aggressively when consistency breaks. Random regeneration without changing the root cause produces different failures, not solutions. Every regeneration costs credits. The correct sequence is: identify the failure mode, fix the root cause in the brief or the anchor frame, then regenerate once to validate the fix before continuing the sequence.

A character who reads as the same person across six shots in a row is a system, not luck. The system is a tight brief, a locked anchor frame, a sequenced Film workflow, and a Canvas that makes it repeatable.

The consistent character does not come from the right single prompt. It comes from treating every shot as a node in a connected system rather than a standalone generation. Brief discipline, visual anchoring, shot sequencing, and workflow automation are each doing part of the work. Remove any one of them and the character drifts. Keep all four and the consistency holds across a full project, not just across the first three shots where the model's initial settings happen to align.