Seedance 2.0 Glossary

When public guides mention character consistency, they usually mean keeping the same face, outfit, silhouette, and overall identity across multiple shots. In practice, that depends on reference hygiene: use a small number of clear images, keep those images visually close to the result you want, and repeat the same identity constraints in each prompt. If consistency breaks, the model is often being asked to carry too many visual variables at once.

Video extension means continuing an existing clip from its last frames rather than generating a brand-new scene. Multi-shot storytelling means designing several shots that still feel like one sequence. These ideas are related but not identical: extension is mostly about temporal continuity, while multi-shot planning is about story structure, pacing, camera changes, and visual anchors that survive from one shot to the next.

Multimodal input means combining text with images, video, and sometimes audio cues in one request. Native audio means the output can include generated sound instead of relying only on silent video. Lip-sync refers to mouth motion aligning with spoken dialogue. These terms matter because they change how you brief the model: a text-only prompt needs more descriptive detail, while a multimodal prompt can delegate part of the control to references, source footage, and audio timing.

2K output refers to the maximum video resolution available in Seedance 2.0, roughly 2048 pixels on the long edge. Public materials describe this as a significant step up from earlier model versions that often capped at 720p or 1080p. In practice, 2K matters most when output will be viewed on larger screens or cropped in post-production. Not all access surfaces may expose the full 2K option at every pricing tier, so verify what your specific platform offers.

Camera motion replication means uploading a reference video clip and telling the model to copy its camera movement — pans, tilts, tracking shots, dolly zooms, orbits, or handheld shake — rather than describing the motion in text. In Seedance 2.0, you tag the reference clip with @video1 and specify its role as camera guidance. This is especially useful when you want a precise movement style that is hard to describe in words, such as a specific Steadicam feel or a particular drone flight path.

Text-to-video (T2V) generates a clip entirely from a written description. Image-to-video (I2V) starts from one or more uploaded images, so the model inherits composition, subject appearance, and color from the source. Most production workflows mix both: use I2V when you need visual fidelity to existing assets (product shots, storyboard frames, character designs) and T2V when you want the model to generate the full scene from imagination. Seedance 2.0 also supports hybrid workflows where images supply the look and text supplies the motion.

References are the source files you upload alongside a text prompt to guide the generation. In Seedance 2.0's All-round Reference mode, each file is tagged with an @mention (e.g., @image1, @video1, @audio1) to specify its role: first frame, character face lock, style palette, camera motion source, or music sync track. Reference hygiene — using a small number of clear, visually consistent files — is one of the biggest factors in output quality. Conflicting references (different lighting, mismatched faces, clashing styles) are a common cause of drift and artifacts.

Identity drift is the visible change in a character's appearance — face shape, hairstyle, clothing details, or body proportions — between frames or across shots within the same project. It typically occurs when the model receives overloaded references, conflicting style anchors, or prompts that introduce too many visual variables at once. The most reliable countermeasure is reference simplification: fewer images, consistent lighting across references, and repeating the same identity-locking phrase in every generation request.

Reference-first architecture is a workflow philosophy where uploaded images and videos drive the generation more than the text description. Instead of writing a 200-word paragraph describing every visual detail, you supply clear reference files and use the text prompt mainly for motion, timing, and context that the images cannot convey. This approach tends to produce more visually consistent results because the model inherits composition, color, and identity directly from the source material rather than interpreting ambiguous language.

A neural asset is a reusable AI representation of a character, product, or scene that has been encoded via reference images so the model can reproduce it consistently across multiple generations. Unlike traditional 3D assets, a neural asset lives in the model's latent space and is invoked by uploading the same reference set with consistent tagging. Building a reliable neural asset means curating 2–5 high-quality reference images that agree on lighting, angle, and costume, then testing them across several generations before committing to a production run.

A canonical character sheet is a multi-angle reference collage — typically front, side, and three-quarter views — used as a persistent identity anchor for a character across shots. The concept borrows from traditional animation model sheets. In Seedance 2.0 workflows, uploading a well-made character sheet as a reference image gives the model a richer spatial understanding of the subject than a single photo can. Best results come from sheets where all angles share the same lighting, background, and outfit.

Prompt leakage happens when style elements, color palettes, or visual motifs from one generation bleed into an unrelated generation, often within the same session or when conflicting references remain active. It is distinct from identity drift because the contamination comes from a previous prompt context rather than from the current reference set. Mitigation strategies include starting a fresh session for unrelated projects, clearing uploaded references between runs, and keeping style anchor images separate per project.

Keyframe injection is the practice of re-uploading reference images at specific moments in a multi-shot workflow to reset the model's visual memory and prevent gradual drift. For example, if you are generating a 5-shot sequence, you might re-inject the original character reference at shot 3 and shot 5 instead of relying on the model to carry identity forward from shot 1 alone. This technique is especially important for longer sequences where cumulative drift would otherwise make the final shot look noticeably different from the first.

The anchor-and-master method is a production workflow where one master shot establishes the definitive look — lighting, color grade, composition, and character appearance — and all subsequent shots are generated using that master as the primary reference. This creates a visual anchor that keeps the entire sequence cohesive. The method works best when the master shot is generated first, reviewed carefully, and only promoted to anchor status after it meets quality standards. Changing the master mid-project usually requires re-generating downstream shots.

SCELA stands for Subject + Context + Effect + Lighting + Action. It is a structured prompt formula designed to produce consistent, predictable results by ensuring every generation request covers the five key dimensions the model needs. Subject defines who or what is in the scene. Context sets the environment and mood. Effect specifies stylistic treatment. Lighting controls the visual atmosphere. Action describes what happens and how the camera moves. Following SCELA reduces the chance of the model filling in gaps with unwanted defaults.

Shot-script format is a timecoded prompt structure where each segment of a video gets its own instruction block — for example, '0–3s: wide establishing shot of city skyline at dusk, 4–8s: cut to medium shot of character walking, 9–13s: close-up on hands opening a letter.' This format is designed for 13–15 second Seedance 2.0 generations and gives the model explicit timing cues for transitions, camera changes, and action beats. It works best when each segment has one clear action and one camera setup.

Content decay describes the rate at which AI-generated video content becomes visually or technically obsolete as newer models improve. A clip that looked impressive when generated by an earlier model may appear stiff or low-fidelity compared to outputs from the next generation. This matters for production planning: teams building libraries of AI video content should factor in re-generation cycles, version their assets, and avoid over-investing in polish for content that will be superseded within months.

Negative prompting — writing instructions like 'no blur,' 'don't show hands,' or 'avoid text' — is a technique from image generation models that does not work reliably in Seedance 2.0. The model may parse the descriptive noun ('blur,' 'hands,' 'text') and actually increase the presence of the unwanted element. The recommended alternative is positive phrasing: instead of 'no blur,' write 'sharp, in-focus, high clarity.' Instead of 'don't show hands,' frame the shot so hands are naturally out of view. This inversion habit is one of the most impactful prompt-writing adjustments for users coming from Stable Diffusion or Midjourney backgrounds.

These are two distinct reference strategies in Seedance 2.0. First/last frame mode pins the opening and closing frames of a video to specific uploaded images, and the model interpolates the motion between them. All-round reference mode uses multiple reference files — images, video clips, audio — as persistent context throughout the generation without restricting specific frames. First/last frame mode gives tighter control over start and end states, making it ideal for product reveals and transitions. All-round reference mode offers more flexibility for complex scenes where identity, style, and motion references all need to coexist.

Video as software is a mental model that treats AI-generated video outputs as updateable, versionable, and iteratable assets rather than one-shot renders. Under this paradigm, a generated clip is not a final deliverable but a draft that can be re-prompted, extended, re-referenced, and remixed — much like source code that gets committed, reviewed, and revised. This shifts production thinking from 'get it right in one take' to 'build, test, iterate,' which better matches the probabilistic nature of generative models and enables faster creative cycles.

A structured approach to describing lighting in video prompts that replaces vague instructions like 'add a light' with precise, reproducible descriptions the model can parse reliably. The three layers are: (1) Source Layer — what creates the light, such as sunset backlight, neon signs, candles, or rim light; (2) Behavior Layer — how light interacts with the scene, including dust diffusion, specular highlights, volumetric fog, and lens flare; (3) Color Tone Layer — the overall palette, for example warm gold base with cool blue highlights, or teal-and-orange contrast. A complete lighting instruction might read: 'sunset backlight dark gold plus core ice-blue self-illumination (source), dust diffusion softening contours plus corroded metal specular highlights (behavior), dark gold warm base plus ice-blue cold-warm clash (tone).' This structure originated in the woodfantasy/Seedance2.0-ShotDesign-Skills project and is designed to give consistent, reproducible lighting across shots.

A production-grade prompt formula that ensures every professional video prompt covers six structural dimensions: (1) Subject and Appearance Details — who or what is in the scene and how they look; (2) Action and Physics Continuity — what happens and how motion stays physically plausible; (3) Scene Environment — the setting, time of day, weather, and spatial context; (4) Visual Style and Physical Lighting — the aesthetic treatment and three-layer lighting setup; (5) Physical Focal Length and Camera Movement — lens choice and camera motion (dolly, crane, orbit, etc.); (6) Native Sound Effects — material-specific sounds and spatial acoustic modifiers. This framework extends the five-element SCELA formula by treating sound as a first-class structural element rather than an afterthought. Professional workflows use six-element assembly to ensure no generation dimension is left to model defaults.

Parameterized visual recipes based on known cinematographic signatures that serve as style anchors for video prompt generation. Using a director preset helps the model converge on a consistent visual language faster than listing individual style attributes. Examples include: Nolan (temporal paradox, IMAX scale, practical effects, deep-focus composites), Villeneuve (vast desolation, geometric framing, minimal dialogue, desaturated palette), Wong Kar-wai (melancholic beauty, step-printed slow motion, saturated color, handheld intimacy), and Kurosawa (compositional power, weather as character, wipe transitions, deep staging). The woodfantasy/Seedance2.0-ShotDesign-Skills project catalogs 28+ presets across Hollywood, Asian cinema, genre styles, and social/commercial categories. You can also combine presets — for example, 'Villeneuve composition with Wong Kar-wai color' — to create hybrid visual identities.

An automatic splitting strategy for videos that exceed the 15-second generation limit. Each segment gets its own self-contained prompt with timestamps starting from 0, a unified style preamble, consistent three-layer lighting across segments, stable handoff frames at segment boundaries (freeze, slow push, or fade), and a shared forbidden-items list. The splitting rule is: ⌈total_duration / 15⌉ segments, with the constraint that the final segment must be at least 8 seconds long. For example, a 40-second video becomes three segments (15s + 15s + 10s), not four segments with a short tail. Narrative arc is distributed across segments: opening, development, climax, and resolution. This approach ensures visual and tonal consistency even when the model generates each segment independently.

A production philosophy that replaces generic filler words — 'masterpiece,' '4K,' '8K,' 'ultra HD,' 'ultra-clear' — with specific physical material descriptions that give the model concrete visual targets. Instead of 'ultra-clear,' write 'Kodak 5219 film stock warmth, brushed aluminum reflections, wet concrete texture.' Instead of 'high quality,' describe the actual surface properties and imperfections you want: film grain type (Fuji Eterna cool shadow), material texture (aged leather, corroded metal), or organic artifacts (lens dust, subtle focus breathing, micro camera shake). Filler words are abstract quality claims the model cannot act on; quality anchors are physical descriptions the model can render. This distinction is one of the most impactful improvements users can make to their prompt-writing practice.

A three-tier progressive strategy for handling recognizable intellectual property in video prompts without triggering platform content filters. Level 1 — Name Replacement: swap the recognizable name for an original descriptive nickname (for example, 'Iron Man' becomes 'Alloy Sentinel'). Level 2 — Feature Modification: replace iconic visual traits like signature color schemes, costume silhouettes, or weapon designs with original alternatives. Level 3 — Category Abstraction: fully abstract the concept so only the narrative role remains, with no visual or nominal connection to the original IP. Each level adds more creative distance from the source material. Always add explicit forbidden items for potential trigger words. Use Level 1 when the character archetype is common enough to avoid filters, escalate to Level 2 or 3 when platform filters still flag the content.

A safety practice for avoiding platform content filter false positives when using camera terminology in video prompts. Bare words like 'Dolly,' 'Crane,' 'Pan,' and 'Dutch' can be misinterpreted by safety filters as proper nouns — person names, brand names, or geographic references — rather than cinematographic instructions. The fix is straightforward: in English prompts, always use full compound phrases such as 'dolly tracking shot,' 'crane shot,' 'Dutch angle tilt,' or 'pan left across the skyline' instead of single capitalized words. In Chinese prompts, use the Chinese camera terms exclusively (推轨镜头, 摇臂镜头, 荷兰角) to avoid English-language filter ambiguity entirely. This practice eliminates a common source of unexpected content flags in otherwise professional prompts.

A structured approach to describing audio in video prompts using material-specific onomatopoeia and spatial acoustic modifiers rather than generic music terms. Sound descriptions are organized into categories: ambient (wind through corridors, rain on metal roof, distant thunder), action (blade cutting air, footsteps on gravel, glass shattering), vocal (whispered dialogue, breath condensing in cold, crowd murmur), and material-based onomatopoeia (silk rustling, metal scraping, wood creaking, water dripping on stone). Spatial modifiers describe the acoustic environment: cathedral reverb, tight room dampening, outdoor open-air, underground echo. This vocabulary works with Seedance 2.0's native audio generation, giving the model physical sound targets instead of vague music requests. Combining material-based sounds with spatial modifiers produces more convincing and immersive audio tracks.