Seedance New: how to write prompts the model actually understands
ByteDance · Updated:
Seedance New is the latest iteration of ByteDance's video model, successor to Seedance 2.0 with experimental refinements. It makes 4–15 second clips up to 2K and accepts up to 9 images, 3 videos, and 3 audio inputs. Noticeably more stable than 2.0 on long videos and consistency scenes. All 2.0 techniques apply in full — this is «2.0+», not a separate architecture.
How New differs from 2.0
New is bleeding edge: the same platform limits (4–15 seconds, up to 2K, 12 files per request), but better long-video generation stability and tighter adherence to complex prompts. On 13–15-second narratives New breaks less often than 2.0.
Consistency Control is experimentally improved — characters drift less between acts, which matters for cinematic scenes with an emotional arc. Documentation is thin, but in practice this model delivers the best results on complex TRY CGI prompts.
Key rating note: as the latest version, all 2.0 techniques (TRY CGI blocks, @-references, per-second timing, identity preservation, micro-acting, Audio Rule, anti-AI anchors) apply in full. Scoring should be maximally strict — the model can handle prompts of any complexity.
- All Seedance 2.0 capabilities in full
- Improved long-video stability (13–15s)
- Less character drift between acts
- Better adherence to complex layered prompts
- Bleeding edge — documentation is limited
TRY CGI prompt structure
Canonical block order: [TITLE & ACT] → LOCATION → REFERENCE ASSIGNMENT → STYLE → STORY → CHARACTERS → SHOT STRUCTURE. One blank line between blocks, a space after every colon.
LOCATION — environment, lighting, weather, key background details. STYLE — visual preset («Ultra-photorealistic 4K live-action cinema», «Gritty film grain»). STORY — what happens in this generation in one or two sentences. CHARACTERS — participants, their current mood, current appearance. SHOT STRUCTURE — the act-based breakdown.
New, like 2.0, internally routes named blocks into different generation layers. On New this effect is amplified: the gap between «wall of text» and a block-structured prompt is wider than on 2.0. If it was ever worth spending time on block structure — it is now.
Timestamp storyboarding and emotional realism
For 13–15-second videos, per-second storyboarding is mandatory. On New it works noticeably more stably than on 2.0: fewer seams and less character drift between acts.
Canonical template: 0–4s [THE ENTRY], 4–10s [THE REVELATION], 10–15s [ACTION RESPONSE]. Each act has fields Action / Emotional Acting / Camera / Lighting / VFX / Audio Rule.
Emotional realism is the main quality multiplier. At least 2–3 micro-signals per act: jaw clenching, pupils dilating, micro-tremor in the eyelids, nostrils flaring, visible heavy breathing, shoulder movement on inhale. Without these even a perfect composition reads as a mask. New renders micro-acting especially convincingly — this is the reason to move from 2.0 to New.
@-references and identity preservation
Full @-syntax works: @image1/@image2/@image3 for characters and scenes, @video1/@video2/@video3 for camera and rhythm cloning, @audio1/@audio2/@audio3 for voice and SFX. Cap is 9 images + 3 videos + 3 audio = up to 12 files.
The critical phrase for any @image of a character — **Strict identity preservation. No morphing or style changes.** On New character drift between acts is noticeably lower than on 2.0, but without this phrase it still appears by second 8–10.
For complex multi-character scenes with two or more @image references, repeat identity preservation for each: «Protagonist (@image1): Strict identity preservation. Antagonist (@image2): Strict identity preservation.» — otherwise the model blends features across characters.
Common mistakes
1. Treating New as «a different model», not 2.0+
All 2.0 rules apply to New in full: TRY CGI blocks, identity preservation, micro-acting, Audio Rule, anti-AI anchors. New is not «re-architected» — it's the same architecture with stability improvements. A blockless prompt on New performs as poorly as on 2.0.
2. Multi-character without identity preservation on each
If the scene has two @image references, identity preservation must be repeated for each: «Hero A (@image1): Strict identity preservation. Hero B (@image2): Strict identity preservation.» Without that the model blends features between characters by mid-clip.
3. Relying on improved stability instead of proper structure
New is more stable than 2.0, but that doesn't mean «write whatever». A solid paragraph instead of block structure, generic emotion words, no micro-acting — all of these produce the same quality drop on New as on 2.0. The stability bonus only applies on top of a properly written prompt.
4. Epic music without the user explicitly asking for it
Default is «No music. Diegetic sound design only.» Universal closing anchor. Music turns a cinematic scene into a trailer, and on 15-second narratives with an emotional arc it kills immersion. If the user wants music, they will ask explicitly.
5. Asking for more than 15 seconds in one prompt
New, like 2.0, is capped at 15 seconds per run. «Make a 30-second clip» will either get truncated or break the pacing. The right path is segmentation: first prompt up to 15s, second segment via @video1 with «Extend @video1 by 15 seconds» plus a new-content description.
Before / after examples
Example 1
Before
long dramatic video where a detective cracks the case
After
LOCATION: Dim 1970s detective office, smoke-stained walls, single desk lamp with cracked green glass shade, evidence photos pinned to a corkboard, half-empty whiskey glass on the desk. REFERENCE ASSIGNMENT: - Detective (@image1): Strict identity preservation. Use this image for exact facial features and wardrobe. No morphing or style changes. STYLE: Ultra-photorealistic 4K cinematic, anamorphic lens flare, gritty film grain, neo-noir color grade with deep shadows and warm key light. The look of Chinatown, not Netflix gloss. STORY: Detective stares at the corkboard, makes the connection that solves the case, and quietly reaches for the phone. CHARACTERS: Mid-50s detective, weathered face, three-day stubble, loosened tie, rolled shirtsleeves. Tired but sharp. SHOT STRUCTURE (15 SEC TOTAL): 0-4s — [THE ENTRY] Action: Detective slowly scans the corkboard, eyes moving photo to photo. Emotional Acting: Quiet focus. Brow furrowed, jaw set, lips slightly parted. Camera: Medium shot, slow handheld drift left to right. Lighting: Warm desk lamp from below, deep shadows on the upper face. 4-10s — [THE REVELATION] Action: Detective stops on one photo, leans in, then sharply pulls back as the realization lands. Emotional Acting: Shock turning to grim certainty. Pupils dilate, jaw clenches, micro-tremor in the eyelids, slow exhale through the nose. Camera: Slow dolly-in to extreme close-up of the eyes, then slight push back. Lighting: Lamp glow sharpens, screen of evidence in the pupils. 10-15s — [ACTION RESPONSE] Action: Detective reaches for the rotary phone, dials without looking, lifts the receiver. Emotional Acting: Cold focus. Brow tension, lips tight, steady breathing. Camera: Tight side angle, handheld, slight motion blur. Audio Rule: Diegetic only — phone dial clicks, glass scraping on the desk, slow inhale. No music. Diegetic sound design only. Grounded realism, gritty film grain, no cartoonish look, no AI aesthetic.
Textbook 15-second TRY CGI prompt that plays to New's strengths: long narrative with three acts, identity preservation on the detective, emotional arc from focus → revelation → cold resolve via explicit micro-acting.
Example 2
Before
two characters in a bar arguing, emotional
After
LOCATION: Smoky neighborhood bar at night, low pendant lights over the booth, neon sign across the window throwing red across one side of the frame. REFERENCE ASSIGNMENT: - Hero A (@image1): Strict identity preservation. Use this image for exact facial features and wardrobe. No morphing or style changes. - Hero B (@image2): Strict identity preservation. Use this image for exact facial features and wardrobe. No morphing or style changes. STYLE: Ultra-photorealistic 4K cinematic, 35mm anamorphic, gritty film grain. The look of a Sicario booth scene, not a sitcom. STORY: Hero A pushes Hero B for an answer; Hero B refuses; the tension cracks but does not break. SHOT STRUCTURE (10 SEC TOTAL): 0-4s — [SETUP] Action: Hero A leans forward across the table, voice low and pressing. Hero B keeps his eyes down on the glass. Emotional Acting: Hero A — controlled pressure, jaw set, eyes locked on Hero B. Hero B — refusal, lips tight, slow breathing, brow heavy. Camera: Two-shot, slow handheld dolly-in. Lighting: Warm key from above, red neon rim on the right. 4-10s — [CLIMAX] Action: Hero B looks up sharply, holds the gaze for two beats, then deliberately picks up his glass and drinks. Emotional Acting: Hero B — quiet defiance. Eyes steady, nostrils flare once, micro-tremor at the corner of the mouth. Hero A — micro-recoil, brief blink, jaw flexes. Camera: Cut to tight shot-reverse-shot on each face. Audio Rule: Diegetic only — glass on wood, low background bar chatter, breathing. No music. No music. Diegetic sound design only. Grounded realism, no AI gloss.
Multi-character scene with two @image references — New holds consistency for both characters more reliably than 2.0. Identity preservation is repeated for each, micro-acting is described separately for each face.
Example 3
Before
extend my 15-second video by another 10 seconds
After
REFERENCE ASSIGNMENT: - Base footage (@video1): Use this as the source. Match exact lighting, color grading, lens, and frame composition of the last frame. Continue camera motion seamlessly. Extend @video1 by 10 seconds. New content begins exactly from the last frame of @video1 — preserve framing, exposure, and color grade. SHOT STRUCTURE (10 SEC NEW CONTENT): 0-4s — [CONTINUATION] Action: The character on screen finishes the action from the previous segment and begins to walk away from camera into the corridor. Camera: Continue the slow tracking from @video1, transition into a wider locked-off shot once the character is mid-corridor. Lighting: Match the existing cool corridor fluorescents from the previous segment. 4-10s — [PAYOFF] Action: Character reaches the far door, hesitates, then pushes it open. Light from the other side spills into the corridor. Emotional Acting: Quiet resolution. Slight pause, shoulders settle, slow exhale. Camera: Locked wide, gentle push-in as the door opens. Audio Rule: Diegetic only — footsteps, door hinge, distant ambient. No music. No music. Diegetic sound design only. Grounded realism, no AI gloss.
Video Extension via @video1 plus a full TRY CGI structure for the new 10 seconds. On New the segment-to-segment seam is smoother than on 2.0 — character drift and color drift are minimized.
