Seedance 2.0 Prompt Template: Copy-Paste Framework for Motion + Camera + Style
A practical prompt framework for Seedance 2.0 with shot language, motion constraints, negatives, and 5 proven templates.

Hey, I’m Dora. A small thing pushed me into this: I kept getting near-misses. The footage looked close to what I wanted, then drifted into a different mood by shot three. I didn’t need more features. I needed a steadier way to talk to the model. So over a few sessions in January–February 2026, I built a Seedance 2.0 prompt template that I could reuse without babysitting every generation.
If you’re still getting familiar with the model before refining prompts, this complete Seedance 2.0 workflow guide covers the broader multimodal creation process and setup. The results I describe below come from my notes; I don’t have saved outputs for all five copy-paste templates.

Prompt anatomy that reduces drift (subject → action → camera → style → constraints)
The biggest lift came from setting a strict order and sticking to it. When I wrote prompts like a sentence, Seedance 2.0 did fine on the first beat, then wandered. When I wrote them like a fill-in card, I saw less drift.
Here’s the five-part spine I use now:
- Subject: Who or what the scene is about, singular if possible.
- Action: What the subject is doing, in plain language.
- Camera: How we see it, shot size + movement + lens cue if needed.
- Style: The look, not the vibe checklist. One anchor reference beats six adjectives.
- Constraints: What to keep fixed, what to exclude, and timing.
Why this order works in my workflow:
- Subject first gives me a center of gravity. If I mention multiple subjects up top, the model can split attention later.
- Action next is the kinetic anchor. It tells the model what must move even if the style shifts.
- Camera then sets framing logic so I don’t leave the lens choice to a vague mood word.
- Style late in the stack adds flavor without hijacking action.
- Constraints last remind me what to check, especially color, lighting, and hands/faces.
A compact Seedance 2.0 prompt template that I copy each time:
Subject: [one person/object, age or material if relevant]
Action: [specific verb phrase, present tense]
Camera: [shot size] + [movement] + [angle], [“wide/normal/telephoto” feel if useful]
Style: [one visual anchor], [lighting], [color treatment]
Constraints: [3–5 exclusions], [tempo], [duration or beat timing], [consistency notes]
An example that held shape across three cuts in my notes:
- Subject: 30s ceramic mug on a workbench, matte white
- Action: Steam rises as a hand slides the mug into frame and pauses
- Camera: Medium close-up, slow dolly-in, eye level, normal lens
- Style: Soft morning window light, subtle film grain, muted palette
- Constraints: No logos, no text overlays, no jump zooms, hold on hand steady for 2s
What changed for me: fewer surprise reframes. Before, I’d ask for “cozy, handheld, morning light” and get a push-in on take one, a shaky pan on take two. The template kept the requested lens behavior steadier without me micromanaging. It did not make every take identical.
Motion + camera vocabulary that actually changes outputs

I stopped using mood words as camera words. “Dynamic” means nothing precise to a lens. Specific motion cues do. When I swapped vague prompts for concrete ones, Seedance 2.0’s motion felt more intentional.
The current Seedance 2.0 text-to-video documentation asks for scene, action, camera movement, and mood in the prompt. Those are descriptive instructions, though, not a guarantee that a generated shot will follow a physical camera path or exact lens setting.
What landed well in my tests:
- Movement words tied to rig metaphors: dolly, track, crane, handheld, gimbal. “Handheld” added micro-wobble; “gimbal” stayed smoother.
- Speed paired with distance: slow, medium, fast, plus a cue such as “slow dolly-in, 1–2 feet.” Even rough numbers helped me describe the move.
- Shot size up front: wide / medium / close gave me more consistent composition and fewer unwanted re-centers.
- Angle with a purpose: eye level for neutral, low angle for presence, high angle for vulnerability or overview.
- Lens cues as buckets: wide (24–28mm feel), normal (35–50mm feel), telephoto (85mm+ feel). I avoid exact millimeters unless I must.
I also found that combining two motion verbs made the model choose chaos. One verb per shot kept things clean. When I needed a compound move (say, pan + dolly), I wrote it as beats: “Start: slow dolly-in. Then: gentle pan right for the final 2 seconds.” Seedance followed that sequence better in my runs than when I jammed both into one clause.
Shot list cheat sheet (wide/medium/close, pan/dolly/handheld)
- Wide: establish space and context. Good for product-in-environment or team scenes. Pair with slow dolly or locked-off. Avoid fast pans unless you want smear.
- Medium: subject + some context. Safe for dialogue and UGC. Handheld here reads personal; gimbal reads polished.
- Close: detail and emotion. Works with tiny push-ins; pans feel jarring. Telephoto cues here help keep background soft.
- Pan: lateral rotate. Use to reveal adjacent info. Keep slow; it compounds motion blur.
- Dolly/Track: physical move toward/away/alongside. Feels cinematic even at low speed. My default for product shots.
- Handheld: slight sway and micro-shake. Great for UGC, risky for text overlays.
I keep that list near my prompt window. It nudges me to pick one clear move instead of a mood paragraph.
Negative prompt checklist (what to ban explicitly)
Bans felt heavy-handed at first, but they saved reshoots. These were the repeat offenders in my runs:
- Visual noise: no text overlays, no watermarks, no floating UI, no lens flares unless specified
- Identity drift: no extra characters, no crowd, no mirrors reflecting other people
- Camera chaos: no snap zooms, no whip pans, no Dutch angles, no jump cuts
- Body artifacts: no extra fingers, no deformed hands, no warped mugs/handles, no melting edges
- Branding: no logos, no labels, no recognizable brands
- Color/grade: no neon lighting, no heavy teal/orange, no cartoon saturation
- Environment: no rain/fog/smoke unless stated, no confetti, no dust particles
- Audio/text: if you’re adding VO in post, ban auto captions
I don’t use all of these every time. I pull 3–5 that matter for the scene. Too many negatives can dull the image. If artifacts persist after two tries, I switch strategy: adjust the subject wording or simplify the camera note rather than stacking more bans.
One API detail matters here: on the current text-to-video route, these exclusions go inside prompt. Its published request fields do not include a separate negative_prompt parameter.

5 copy-paste templates (UGC, product ad, cinematic, talking head, montage)
These aren’t magic: they’re starting points. I paste one in, fill the brackets, and keep the rest of my brain for timing and music. The durations below fit the current Seedance 2.0 4–15-second generation range; choose duration explicitly in the API request rather than expecting the number inside the prompt to set it.
- UGC (phone-in-hand feel)
Subject: [person, age range, setting; if using a supplied portrait, refer to it as @Image 1]
Action: [speaks casually about X while doing Y]
Camera: Medium, handheld phone perspective, slight sway, eye level, normal lens feel
Style: Natural indoor light, ungraded look, light motion blur
Constraints: No captions, no snap zooms, keep hands natural, 8–10s, keep background simple
- Product ad (clean and steady)
Subject: [product name/material/color; if using image-to-video, describe the product in the supplied start image]
Action: [rotates slowly / slides into frame / subtle hero move]
Camera: Close-up to medium close-up, slow dolly-in, locked horizon, normal-to-tele feel
Style: Soft key light + gentle rim, neutral color grade, light film grain
Constraints: No logos/labels, no flares, hold final frame 2s, 6–8s total
- Cinematic (mood-first without losing control)
Subject: [character or place]
Action: [specific beat, waits, turns, breathes, steps into light]
Camera: Wide establishing for 2s then slow push to medium, gimbal-smooth, eye level
Style: [single anchor reference, e.g., “overcast natural light, muted blues”]
Constraints: No Dutch angles, no crowd, no neon, maintain overcast look, 10–12s
- Talking head (stable and legible)
Subject: [speaker description; if using a supplied portrait, refer to it as @Image 1]
Action: [delivers one clear line]
Camera: Medium close-up, locked tripod or very subtle dolly-in, eye level
Style: Soft key from 45°, clean background separation, neutral grade
Constraints: No auto captions, no whip pans, skin tones natural, 12–15s, keep eyeline centered
- Montage (quick beats without chaos)
Subject: [theme, e.g., “morning coffee ritual”]
Action: Beat 1 [wide context], Beat 2 [hands close-up], Beat 3 [steam detail], Beat 4 [sip]
Camera: Each beat about 2s, clear shot size per beat, no compound moves; transitions by cut
Style: Consistent light and palette across beats
Constraints: No text overlays, no speed ramps, keep tempo steady, 8–10s total
Little note from testing: when I want a crisp product edge, I swap “handheld” for “dolly” even in UGC. It looks a hair less authentic but prints cleaner for overlays later.
If you want vocabulary refreshers, the StudioBinder shot sizes guide is handy, and their camera movement overview helps me choose a precise movement word.

Make Each Template Runnable
I keep the prompt text separate from the API settings. Here is the run card I’d put beside the five templates. “Check the result” means what to inspect after generation; I don’t have saved, comparable output clips for all five templates, so these are not claimed test results.
| Template | Input material and route | Set in the request | Check the result |
|---|---|---|---|
| UGC | Prompt alone, or one approved portrait in reference_images ontext-to-video | aspect_ratio: “9:16”, duration: 8–10, resolution: “720p” | Same person, intended sway, readable hands; listen for unwanted speech or captions |
| Product ad | A clean product still as image onimage-to-video | duration: 6–8, resolution: “720p”; choose aspect ratio for the placement | Product shape and material, edge stability, final hold |
| Cinematic | Prompt alone on text-to-video | aspect_ratio: “16:9”, duration: 10–12, resolution: “720p” | Wide-to-medium progression, one coherent move, stable light |
| Talking head | Prompt alone, or one approved portrait in reference_images on text-to-video | aspect_ratio: “9:16”, duration: 12–15, resolution: “720p” | Face and eyeline consistency; verify spoken words and lip sync rather than assuming exact delivery |
| Montage | Prompt alone on text-to-video | aspect_ratio: “16:9”, duration: 8–10, resolution: “720p” | Beat order and cut timing; be ready to assemble separate shots if one generation misses a beat |
For UGC or talking head, put reference image URLs in upload order and name their jobs in the prompt: “@Image 1 is the speaker’s appearance; keep that person consistent.” The Seedance reference documentation uses @Image 1, @Video 1, and @Audio 1 in that order-based form. A file name by itself does not tell the model which role you want the asset to play.
If I already have footage and want to change only its look, I use the video-edit route, which takes the existing clip as video. I write what must stay fixed and what should change. I don’t send the old clip to text-to-video and expect it to preserve the edit.
Here is the failure-and-revision pair from the opening example, without inventing a new render:
- Before: “Cozy, handheld, morning light.” In my notes, one take pushed in and another gave me a shaky pan.
- Revised: “Subject: matte white ceramic mug on a workbench. Action: a hand slides it into frame; steam rises. Camera: medium close-up, slow dolly-in, eye level. Style: soft morning window light, muted palette. Constraints: no handheld shake, no jump zoom, hold the final frame for 2 seconds.”
The revision says which move I actually want. My original notes say the structured mug prompt held its shape across three cuts; they do not provide downloadable clips or a measured success rate for the five templates above.
Decision rules—when to re-prompt vs change reference
When a run misses, I try not to flail. I run this tiny decision tree:
- If framing is wrong but the action is right: re-prompt. Tighten Camera first (shot size + one movement). Keep Subject and Action identical.
- If motion feels off (too wobbly/speedy): re-prompt. Swap “handheld” ↔ “gimbal,” and set a speed. Don’t touch Style yet.
- If style or color drifts while motion and framing are fine: re-prompt. Replace the Style line with a single, stronger anchor and remove extra adjectives.
- If the subject keeps mutating (extra people, changing props) after two re-prompts: change reference if this route has a reference input. Simplify the Subject. Fewer descriptors, one noun.
- If artifacts repeat (hands, labels, weird flares) across three tries: change constraints or the shot plan. Sometimes a close-up is fighting the model: step back to a medium.
Time-wise: I give myself two fast re-prompts (under 5 minutes total). If I’m still nudging the same error, I change the reference line or the shot choice. This kept me from spending an hour sanding one bad idea.
Why it matters in my workflow: the earliest clear instruction gives me a better starting point. If I’ve defined the wrong subject or action there, editing a few style words later rarely fixes the shot.
One last field note: shorter prompts with this structure beat long, poetic ones by a mile in my sessions. My best takes were under 60 words plus constraints.
Previous posts:
/filters:quality(82)/media/images/f34aab8c-40bf-46f1-90c3-f695e4cc62ef.webp)
/filters:quality(82)/media/images/6556391c-8337-4b77-b318-964981612649.webp)
/filters:quality(82)/media/images/1773962750383987480_n3hqzHRZ.webp)
/filters:quality(82)/media/images/c1764ab0-046e-46bf-a1dd-af23e8490929.webp)