14 Seedance 2.0 Prompt Examples With Existing Videos

Use these Seedance 2.0 examples to study reference images, camera directions, materials, and sound. Whether you are searching for copyable Seedance 2 prompts or learning how to use Seedance 2 for the first time, start by matching the shot structure to your scene: a focused continuous take gives you fewer transitions to check, while ordered multi-shot prompts let you request cuts inside one 4–15-second clip.

This gallery pairs fourteen prompt examples with existing videos from ChatCut’s Prompt Library, alongside a complete walkthrough of Seedance 2.0 modes, reference limits, and credit costs. The videos illustrate particular outputs, not a guarantee that a new request will reproduce the same result.

Last verified: 2026-11-07. The capability and pricing facts below come from official ByteDance, Dreamina, and BytePlus documentation and ChatCut’s current documentation. Existing gallery media is preserved; no new generation test was performed for this refresh. This page also folds in our Seedance 2.0 tutorial, quick start, reference formula, and troubleshooting table.

What “production-grade” means for a Seedance 2.0 prompt

The three checks again, with more teeth.

Choose the shot structure deliberately. A continuous take with one camera move gives you fewer transitions to review. Seedance 2.0 can also generate multiple shots in one request: describe Shot 1, Shot 2, and so on, within the 4–15-second duration. Inspect the resulting cut order and continuity instead of treating exact timing or shot count as guaranteed.

Named nouns over vibes adjectives. “Cinematic,” “ethereal,” “premium”: these words describe a feeling, not a picture, so the model has to guess. “Volumetric overhead beam lighting,” “ray-traced reflections on polished floor,” and “razor-sharp edge with grazing rim light” specify concrete lighting and surface details instead of leaving the look open to interpretation. The first style of prompt produces generic output; the second gives you a specific visual target you can verify.

Reusable as-is. If the output needs frame-by-frame retouching, fingers fixed, identity stitched, the prompt didn’t work. It might be a beautiful clip; it isn’t a production asset. The existing videos below show particular outputs. For your own project, watch the full result and check details, sound, and continuity before approving it as a final asset.

If you want the formula side of this, how to assign roles to images, videos and audio, jump to the reference formula below the gallery.

How to use Seedance 2: modes and limits

Seedance 2.0 is a multimodal audio-video model from ByteDance Seed that you run through web services, APIs, or integrated editors rather than a standalone local installer. Understanding how to use Seedance 2 starts with picking an access route, choosing the right input mode for your assets, and checking the resolution and credit rules before you generate.

Where can you use Seedance 2.0?

• Official consumer web app (Dreamina). Dreamina advertises global browser access to Dreamina Seedance 2.0 with up to 12 combined references (capped at 9 images, 3 videos, and 3 audio files) and daily free credits. Individual account or regional eligibility and stable subscription pricing were not tested; check the live Dreamina interface for your account’s current offer.
• Official developer console and API (BytePlus ModelArk). BytePlus provides the Create video generation task API, the Get task API, and an online Model Playground for dreamina-seedance-2-0-260128, Seedance 2.0 Fast, and Seedance 2.0 mini. Enabling the model requires meeting BytePlus activation conditions, such as an account balance above $30 or an eligible AI Savings Plan or resource pack.
• Integrated timeline editor (ChatCut). ChatCut includes Seedance 2.0, Seedance 2.0 Fast, Seedance 2.0 mini, and Seedance 2.5 inside Video Gen on the web app, on ChatCut Desktop (macOS 13+ on Apple silicon or Intel; Windows 10+ x64), and through the ChatCut Agent Plugin for Claude Code, Codex, and WorkBuddy (Product Facts). Generated clips land in My Assets so you can cut, caption, and mix them on a multi-track timeline.
• Third-party web hubs. Services such as OpenArt, WaveSpeed, and KreadoAI also expose Seedance 2.0 inside their own web interfaces and credit systems. Treat them as third-party integrations with platform-specific menus rather than official ByteDance properties.

Choose text, frame, or reference mode

Every Seedance 2.0 interface maps to four core workflows documented in the BytePlus Seedance 2.0 prompt guide and ChatCut’s Model Capability Matrix:

1. Text-to-video (invent a shot). Use text alone when you want to establish a scene, mood, or b-roll shot without locking an existing product photo or character face.
2. First frame or first + last frame (animate or bridge stills). Upload a First Frame to lock opening composition, product geometry, and lighting, then describe the motion that follows. Add a Last Frame when the shot must land on a specific ending pose or frame. In ChatCut, a last frame on Seedance 2.0 requires a first frame, and frame slots cannot be combined with multimodal reference slots in the same call.
3. Multimodal reference mode (borrow motion, lock identity, or guide audio). Attach up to 9 images, 3 reference videos, and 3 audio clips (up to 9 attached files total in ChatCut’s Video Gen; each video or audio file must be 2–15 seconds, with a 15-second combined cap per media type and video files up to 200 MB per the BytePlus tutorial). Audio references require at least one image or video reference. State explicitly what each file controls (@image_1 for character outfit, @video_1 for camera movement only).
4. Source-video editing vs. motion reference. A reference video in Seedance 2.0 guides camera work, motion, or pacing for a newly generated clip; it is not an in-place editor for the uploaded video. In ChatCut, dedicated Edit video and Extend video input modes require switching the model selector to Seedance 2.5 (AI Video Generation).

If you want a guided reference-video workflow without configuring every slot by hand, open the Hanging Orbit Skill and supply an approved adult character image.

Set duration, aspect ratio, resolution, and budget

• Duration and aspect ratio. Full Seedance 2.0 generates 4–15-second clips across six fixed aspect ratios (16:9, 9:16, 4:3, 3:4, 21:9, and 1:1) plus adaptive where the mode supports it. Starting with a 4–6-second draft at 480p or 720p is a practical editorial habit to verify framing and subject stability before paying for a 15-second take.
• Resolution tiers vs. export resolution. Full Seedance 2.0 supports 480p, 720p, 1080p, and native 4K (10-bit HEVC in ChatCut, which takes longer and costs about twice as much as 1080p). Seedance 2.0 Fast and Seedance 2.0 mini output at 480p and 720p for lower-cost drafts. Keep generation resolution separate from timeline delivery: ChatCut Desktop renders local exports up to 4K, whereas cloud exports go up to 1080p (Web vs Desktop Export).
• Pricing and credit rates. On the BytePlus ModelArk pricing page and billing rules, API charges multiply token consumption by the model’s unit rate, increasing with resolution, duration, and video inputs (failed tasks are not billed). In ChatCut, video generation requires a paid plan (starting at $25/month for 100 credits or $252/year billed annually; 1 credit has a $0.25 face value). Per ChatCut’s Credits policy for current-pricing accounts (earlier subscriptions may show lower rates while active):
  • Seedance 2.0: 0.336 credits/sec at 480p (shown as starting at 0.34/sec in the Upgrade dialog), 0.72 credits/sec at 720p (about 3.6 credits for a 5-second clip before rounding up to the nearest 0.01 credit), 1.584 credits/sec at 1080p, and roughly double the 1080p rate for native 4K.
  • Seedance 2.0 Fast: 0.264 credits/sec at 480p and 0.576 credits/sec at 720p.
  • Seedance 2.0 mini: 0.168 credits/sec at 480p and 0.36 credits/sec at 720p.
• What costs nothing in ChatCut. Uploading media, manual timeline editing, transcription, standard caption editing, and exporting (MP4/WebM video, MP3 audio, subtitles, and XML) never use ChatCut credits (Free and Pro). When you run ChatCut through Claude Code or Codex (plugin or MCP), the agent’s editing work (cutting, reordering, removing pauses and fillers, captions, B-roll placement, and motion graphics the agent writes) is covered by your existing Claude or ChatGPT subscription and uses no ChatCut credits; ChatCut credits apply only when calling a ChatCut-hosted generation service.

Quick start: your first Seedance 2.0 video

If you’ve never run Seedance 2.0, start small and change one thing at a time.

1. Select the model and input mode. In ChatCut, switch the AI composer to Video Gen, click the model selector below the composer to choose Seedance 2.0, and open the Video settings slider icon (AI Video Generation). Under Input, pick text, frames, or references. (In Dreamina, open AI Video, choose Dreamina Seedance 2.0, and use the + upload slots.)
2. Set aspect ratio, resolution, and duration, then upload your reference or drag compatible media from My Assets (Generation References). One image plus a short brief is enough for simple motion; add a video reference when you want a specific camera move or blocking, and an audio reference when sound matters.
3. Write four things: what’s visible, what changes, in what order, and how the camera moves. Check the credit estimate beside Send message before submitting.
4. Generate one short clip in My Assets, watch it all the way through with sound, and drag approved takes to the timeline. If something misses, find the single biggest problem (subject, camera, pacing, or mood) and change only the instruction that controls it using the troubleshooting table.

A one-image example, written as ordered actions rather than exact timings:

14 Seedance 2.0 prompt examples with videos

The order is curatorial, not chronological. Logo work first because it’s the safest starting point. Product and commercial second because that’s where most readers arrive from. Character-driven third, where the model’s documented weaknesses bite hardest and the prompts have to defend against them. Structural and effects work last, where Seedance 2.0 does things earlier video models couldn’t.

Before copying a card below, distinguish between a preset shorthand and a complete manual brief. Each original one-line prompt comes from a ChatCut preset whose hidden system instructions supply lighting, camera moves, and guardrails. Copying text alone never attaches media files. To run any prompt outside its preset: replace every @image:Name token with an uploaded file mapped to that subject slot, fill every {{field:hint}} with your own scene details, and delete instructional notes such as {{field: please avoid uploading real people}} so they do not become scene text. Finally, select the target duration, aspect ratio, and resolution in your interface, and state what each attached image, video, or audio reference controls.

1. Cinematic Logo Animation

Standalone manual brief (newly authored for direct Video Gen input; untested, not the preset prompt that produced the video below):

Settings & inputs: 1 image reference (@image:logo); requests 6s, 16:9; original resolution and seed not found.

The visible prompt is short on purpose. What earns the result is everything you don’t see: the system-level instructions the preset wraps around the user’s logo, which specify the lighting language. Volumetric overhead beam, ray-traced floor reflection, grazing rim light, razor-sharp edge clarity. The trick to read in this one is that nothing in the prompt describes the logo. Logos animate themselves when you give them light. Describing the shape adds noise to a model that already knows what a logo looks like.

A logo reveal is a focused place to start. Check the logo shape, edges, and lettering throughout the clip; keeping people out of the shot does not guarantee detail preservation.

2. Liquid Logo Animation

Settings & inputs: 1 image reference (@image:logo); requests 6s, 16:9; original resolution and seed not found.

The change from prompt #1 to this one is a single compound noun: liquid-particle. Same logo input, same duration, same aspect ratio, and Seedance 2.0 aims for a fluid assembly rather than a rigid light reveal because the prompt names a concrete physical material.

When you want a stylized reveal, name a tangible substrate (water, mercury, oil, or gel droplets) rather than vague mood adjectives like “magical” or “ethereal.” Concrete material nouns specify how surfaces should catch light and coalesce into the final mark; after generation, inspect whether the particle edges settle cleanly and the final logo lettering stays legible.

3. Glitch Logo Animation

Settings & inputs: 1 image reference (@image:chatcut-logo); requests 4s, 16:9; original resolution and seed not found.

Glitch is the camera move you don’t write. The user prompt names the effect, RGB-glitch reveal, and the system instructions name the artifact vocabulary: chromatic aberration, RGB channel split, horizontal scanline displacement, datamosh frame interpolation. What’s missing from the prompt is any camera direction at all. That’s deliberate. When the effect is the subject, you let the model invent the camera; if you tell it to dolly and orbit while it’s also trying to glitch, it splits attention and lands neither.

Read this one as a counterweight to prompts 1 and 2. Cinematic uses cinematography vocabulary because the effect is light; this one uses artifact vocabulary because the effect is data corruption. Use the noun family the effect lives inside. Don’t borrow vocabulary from the wrong domain.

4. Product Launch Ad

Standalone manual brief (newly authored for direct Video Gen input; untested, not the preset prompt that produced the video below):

Settings & inputs: 1 product image (@image:Sneakers); requests 15s, 16:9; original resolution and seed not found.

This preset describes several named camera positions. The system instructions describe several positions: extreme macro with grazing light, ultra-smooth dolly push from mid-shot to three-quarter, high-speed 360-degree orbital around the product, low-angle power shot from floor level, lateral tracking sweep across the full profile, top-down overhead descent into a bird’s-eye frame, and a final pullback to a wide centered hero.

Use concrete directions such as macro, dolly push, orbital, low angle, lateral tracking, and bird’s-eye. Naming the intended order makes the brief clearer; review whether the actual output follows it.

If you’re working on product ads or browsing the Product ads & UGC Skill collection, this is the template to copy first. The visible user prompt is one line; the heavy lifting sits in the system layer the preset ships around your one-line input.

5. Premium Hair Care Ad

Settings & inputs: 1 product image (@image:Shampoo); requests 10–15s, 16:9; actual duration and resolution not found.

This is the inverse of the Product Launch Ad camera grammar. When the product itself moves, a shampoo pour, a fragrance mist, a serum drip, you lock the camera and give the verbs to the product. The system prompt here calls for a slow tilt and a held wide; that’s intentional restraint, not laziness.

The lesson reads cleanly. Not every product prompt needs six camera positions. If the substance is doing the visual work, the camera should be the audience watching the substance. Sneakers don’t move; the camera must. Shampoo moves; the camera must not. Pick a side per product.

6. ASMR Product Video

Standalone manual brief (newly authored for direct Video Gen input; untested, not the preset prompt that produced the video below):

Settings & inputs: 1 product image (@image:Product); requests 10–15s, 9:16; actual duration and resolution not found.

This is the prompt where Seedance 2.0’s audio generation pulls its weight. The model produces synchronized SFX with the video: peel, snap, crinkle, the precise consonants of a wrap being opened. ByteDance’s launch notes describe this as unified multimodal audio-video generation, and this example asks for sound and picture together. The audio doesn’t come from a separate stem you composite in. It arrives baked into the file.

The vocabulary that earns the result is foley vocabulary, not cinematography vocabulary. Peel, snap, crinkle, micro-tap, soft impact, paper friction, plastic wrinkle. If you write a Seedance 2.0 audio-led prompt in cinematography nouns, you’ll get a quiet handsome video. If you write it in foley nouns, you’ll get a sound-design clip with picture attached. That’s the right way around for ASMR.

7. My Product is in Times Square!

Settings & inputs: 1 product image (@image:Product); requests 10–15s, 16:9; actual duration and resolution not found.

Naming a recognizable location such as a Times Square billboard sets both the display surface and the surrounding architecture, street lighting, and crowd motion in a few words, whereas asking for “a screen” leaves the environment unspecified.

Other specific settings you can request the same way include a side-of-building projection in Tokyo, a stadium jumbotron, a Las Vegas Sphere wrap, or a subway-platform digital display. Name the intended venue clearly, then inspect whether the product perspective, screen borders, and brand details remain accurate in the generated output; this gallery does not establish a measured advantage for named venues or predict moderation outcomes.

8. Transformer is Here!

Settings & inputs: 1 vehicle image (@image:car); requests 15s, 16:9; original resolution and seed not found.

Describe the two endpoints and the transformation between them. This example asks for a car to become a mechanical creature. Use assets and subjects you have permission to use, and follow the provider’s reference and content rules.

Fifteen seconds is the upper end of full Seedance 2.0’s base-generation range, not a measured minimum for every morph. Give the transformation enough screen time for your scene, then inspect the intermediate geometry and product shape.

9. My Character in a Hero Scene

Settings & inputs: Text character field (avoid real-person uploads); requests 15s; original ratio and resolution not found.

This example narrows the brief to one character, one shot, and one action. That reduces the number of cuts to review, but identity and anatomy can still drift within a continuous take. Use an approved reference and check the whole clip.

The placeholders request a character description and advise against uploading real-person photos for this preset. Reference eligibility depends on the provider’s current rules and asset route; ordinary face-bearing uploads, authorized portrait assets, and eligible trusted outputs are not interchangeable. Do not treat a stylized character as a guarantee of consistency.

10. Comic to Live-Action Film

Settings & inputs: 1 comic panel image (@image:Comic); requests 10–15s, 16:9; actual duration and resolution not found.

Uploading a comic panel supplies a concrete visual reference for line weight, color palette, contrast, and framing rather than relying on broad style labels like “cel-shaded” or “noir comic-book” alone. State in your prompt which elements from the panel should stay fixed (character costume, lighting mood, camera angle) and how the scene should move as live action.

If you’re building AI filmmaking workflows, this pattern is worth testing whenever a visual look is easier to show than to describe in prose. Reusable visual references can reduce how often you need to re-explain the palette; inspect each generated shot to confirm how closely it preserves the reference look.

11. Cat & Dog Podcast Skit

Settings & inputs: 1 image (@image:Pets) plus two dialogue fields; requests 15s, 16:9; original resolution not found.

This example uses two stylized animals for a short dialogue scene. Check each character’s appearance and mouth motion throughout; stylization does not remove the need to review identity and continuity.

The two-line dialogue slot is also doing structural work. By naming Luna and Buddy and giving each a single sentence, the prompt tells the model where the dialogue beats sit in the 15 seconds. Review mouth motion against the generated speech; this example does not establish a lip-sync advantage over photoreal subjects.

12. Storyboard to Film

Settings & inputs: 1 nine-panel image (@image:Storyboard); requests 10–15s, 16:9; actual duration and resolution not found.

This preset uploads a nine-panel storyboard sheet as one reference image and asks Seedance 2.0 to follow the panels in order. It is one way to guide several shots, not the only way: the official guide also uses text to request cuts and shot sequences within a single generation.

Giving each storyboard panel a distinct shot size, angle, or action (such as wide, medium, close-up, reaction, and environment cut) makes the intended sequence clearer than repeating near-identical frames. Watch the full generated clip alongside your sheet to check whether any beats were merged or skipped, as panel order and shot count are not guaranteed.

13. 3 Frames → One-Shot

Settings & inputs: 3 reference images (@image:Frame 1–3); requests 15s, 16:9; original resolution and seed not found.

Upload three images as references for a requested start, middle, and end. This is reference-guided prompting, not a dedicated three-keyframe timing control or proof that every supplied frame will appear exactly. The “no cuts, no crossfades” constraint at the end of the prompt is load-bearing. Without it, the model occasionally inserts a soft transition between frames and breaks the one-shot illusion.

The three frames work best when they imply a single physical action. Person walking toward a doorway, opening the door, stepping through, all from the same camera angle. The interpolation has somewhere to go. Three frames that imply separate actions, like person at desk, then person in car, then person on beach, produce surreal teleport sequences; that can be intentional, but it isn’t what this preset is tuned for.

14. Time Freeze Effect

Settings & inputs: 1 scenario text field; requests 15s, 16:9; original resolution and seed not found.

Time freeze and bullet-time are concise cinematography terms for moving an orbital or tracking camera through a stationary scene. Rather than relying on the term alone, pair it with explicit visual details: frozen subjects, suspended water droplets or debris in mid-air, an orbital camera move, and zero motion blur on background elements.

The system instructions for this preset spell out those constraints (absolute stasis instead of slow motion, suspended environmental particles, and sharp edges across the frame) so the brief clearly separates a frozen moment from a slow-motion shot. Inspect the generated clip to verify whether background characters and particles actually hold still while the camera moves.

Patterns to reuse from the examples

Match the structure to the scene. Some examples request a continuous take; product and storyboard examples can call for several shots. Seedance 2.0 supports both. Use ordered shot descriptions when cuts matter, and check their order and continuity after generation.

Name the camera and action early. Put the intended shot type, movement, subject, and lighting where a reader can understand the brief quickly. That is a writing recommendation, not a measured model-attention rule.

Use concrete visual language. Macro, dolly, orbital, tracking, low angle, and top-down describe different camera intentions. Specify the one your shot needs rather than stacking several incompatible movements.

Use references deliberately. A reference can guide subject, style, or composition. Select the right input role and stay within the model’s count and combination limits; visual preservation still needs review.

State the constraints that matter. If you need no cuts, no added text, or a stable product shape, say so and check the full result. A constraint is a request, not a preservation guarantee.

A reference formula you can reuse

Most weak Seedance prompts attach several files and never say what each one is for. Give every asset a role:

Then write the prompt in this order: asset and role → action → order of beats → camera → sound → constraints.

Weak: “Use these three files to make a cool chase video.”

Role-assigned:

A product example in the same shape:

Variations worth knowing:

• Lock a character: attach front and side photos plus an outfit reference, and name the details that must stay (hair color, jacket, glasses). Check identity in every shot.
• Lock a product: add references for the silhouette, the profile and close-ups of texture or hardware. Check the geometry and branding frame by frame.
• Borrow a camera move: say which part of the video reference to borrow (“the dolly push and timing only”) and describe the new subject and scene separately.
• Separate motion and scene: a character image, an environment image and a video reference for the performance, each named.
• Cut to music: attach the track as an audio reference with your images, and describe where beats should trigger cuts, text or scale changes. Expect to fine-tune sync in the editor.

For more examples of turning still photos into controlled motion across models, see our guide to image-to-video prompts. If you are planning recurring episodic or faceless channel production, read our Seedance 2.0 YouTube channel workflow.

Seedance 2.0 facts that change how you write prompts in 2026

Full Seedance 2.0 generates 4–15-second clips and accepts text, images, video, and audio through its documented input modes. It supports six fixed aspect ratios (16:9, 9:16, 4:3, 3:4, 21:9, and 1:1) plus adaptive where the mode permits it. Reference mode allows up to nine images, three videos, and three audio clips within the combined request limits; audio references need an image or video reference.

Official full-model output includes 480p, 720p, 1080p, and 4K, and ChatCut’s model matrix lists native 4K for Seedance 2.0 as of November 1, 2026. Fast and Mini support 480p and 720p. Generation resolution, any upscaling, and final editor export are separate facts.

The provider enables synchronized audio by default and documents an audio-off control; the inspected ChatCut route keeps audio enabled. Multi-shot content still has to fit the total clip duration. For longer sequences, assemble approved clips or use a supported edit/extension workflow with its own input limits. The provider documents editing and extending through video references; in ChatCut, separate Edit and Extend controls are documented for Seedance 2.5, not 2.0.

For Seedance 2.0, describe ordered shots rather than promising exact second-level timing. Seedance 2.5’s timestamp-following guidance is a separate model fact. Moderation, asset eligibility, request limits, and reference availability can cause failures; inspect the actual task error instead of assuming a particular word or subject caused it. See the dated source table above.

How to actually run these prompts in 2026

Choose an available provider route and check its current input roles and limits. The short prompt shown in a gallery card may rely on preset instructions covering camera, lighting, constraints, and reference routing. Copying the line into another wrapper does not reproduce all of that context. For step-by-step setup across modes and providers, see our how to use Seedance 2 walkthrough above.

In ChatCut, use the Prompt Library to inspect an existing preset, add your own eligible reference, and review the model settings and estimate before submitting, or configure the model manually following the AI Video Generation guide through an available AI video generation route. You can also see end-to-end commercial examples in AI-Generated Ads and check the model matrix for current options.

You can generate or import a still first, approve its composition, then use it as the video reference. Place the approved clip in the project and review it against the rest of the edit. Use the credits policy and the current estimate rather than an old flat example: it lists Seedance 2.0 at 720p at 0.72 credits per second, so a five-second clip is about 3.6 credits before rounding for current-pricing accounts; earlier subscriptions may differ. Video generation requires a paid plan, while manual timeline editing, transcription, captions, and exporting use no ChatCut credits.

When the result isn’t what you pictured

Change the instruction that controls the problem, and leave everything else alone.

Two more checks. If you asked for one continuous take, watch for hidden cuts or speed changes anyway; “no cuts” is a request, not a guarantee. And if a task fails, read the actual error: input limits, file eligibility and moderation all produce different messages.