Seedance 2.0 Prompts vs. References: What to Write, What to Upload, and How to Keep a Shot on Brief

TL;DR

Use the prompt to direct what happens: the subject’s action, timing, camera movement, and any changes you want. Use a reference video when you need to anchor a particular movement, performance, or visual continuity. Start with text-to-video when the scene can be invented from your description. Keep each reference’s job explicit, and revise one instruction at a time after reviewing the result. Seedance prompting guides describe both text-only input and reference-video workflows, but this guide is about how to plan the brief—not a guarantee that a model will reproduce it exactly. (https://blog.siray.ai/seedance-2-prompt-guide/)

Key takeaways

• Put actions and shot behavior in the prompt; use a reference to show the movement or visual information you need anchored.
• Give each reference one clear job: for example, “match this camera move” or “preserve this performance.”
• Keep the first pass focused: one subject, one main action, one camera direction, and only essential constraints.
• If a result misses, classify the miss as appearance, action, or camera before editing the prompt.
• When using a reference clip, trim it to the relevant segment rather than asking the model to infer your target from unrelated footage. (https://blog.siray.ai/seedance-2-prompt-guide/)

Give the prompt and reference different jobs

A useful Seedance 2.0 prompt guide begins with a simple distinction: the prompt is direction; the reference is an anchor. Your words tell the model what should happen. A reference clip can help show the movement, camera path, or performance you want it to follow. Seedance guidance describes text-only generation as working from a written description without a concrete reference, and recommends reference-to-video when a specific camera move, action sequence, or performance needs to match a vision. (https://blog.siray.ai/seedance-2-prompt-guide/)

What belongs in the prompt?

Describe what changes over time. “A ceramic mug on a table” gives a subject and setting. “Steam rises as the camera slowly arcs left” adds visible action and a camera instruction. If the shot depends on a sequence, say what happens first and what happens next. If facial expression or sound matters, name it rather than assuming the model will infer it. The Seedance guide specifically recommends spelling out motion, timing, expression, and sound for complex scenes. (https://blog.siray.ai/seedance-2-prompt-guide/)

A practical prompt order is subject → action → camera → style → constraints. Another guide recommends subject, action or motion, camera movement, style, lighting, and quality modifiers. The order is a planning aid, not a requirement to fill every slot. Keep the prompt to a few important elements so that its instructions do not compete. (https://wavespeed.ai/blog/posts/blog-seedance-2-0-prompt-template/) (https://sagnikbhattacharya.com/blog/better-prompts-seedance)

What belongs in a reference?

Use a reference clip when the target is easier to show than describe: a particular camera move, action sequence, or performance. The Siray guide describes downloading a source clip, trimming it to the relevant segment, and using that segment as the generation reference. That gives your reference a defined purpose and avoids asking a model to interpret a long clip when only one part matters. (https://blog.siray.ai/seedance-2-prompt-guide/)

A reference is not a complete creative brief. It may show the movement you want but also contain a background, subject, or composition you do not want. State what the reference should demonstrate, then use the prompt to direct what should change. Do not assume every video tool accepts the same reference types; check the input options of the specific tool you are using.

Choose text-to-video or reference-to-video by the risk of guessing

Start with text-to-video if the scene can be invented from your description and there is no exact source to match. A VideoWeb guide recommends starting with text-to-video unless strict character continuity is needed. Reference-to-video is the better test when the camera move, action sequence, or performance is central to the brief. (https://videoweb.ai/blog/detail/Seedance-2-0-Video-Generation-Guide-Tutorial-Prompts-578fb91b8f46/) (https://blog.siray.ai/seedance-2-prompt-guide/)

Use this decision test: Would a reviewer reject the shot because it invented the wrong movement or performance? If yes, consider a reference clip when your tool supports one. Would they reject it because the action itself is wrong? Make the action and timing more explicit in the prompt. If both appearance and movement matter, make the reference’s role specific and describe the action separately.

Treat the first result as a test, not a promise. Watch it and identify the most important miss. If the action is right but the camera is wrong, revise the camera instruction—not the lighting, setting, and subject all at once. A focused edit makes it easier to tell whether your change helped.

Keep the prompt and reference from contradicting each other

A brief becomes hard to follow when its directions conflict. For example, a reference shows a fixed overhead view while the prompt asks for a fast orbit. Or the reference shows a slow, deliberate gesture while the prompt asks for a sudden movement. Resolve that conflict before generation: choose which instruction governs, then state it plainly.

Write down the reference’s role in a short phrase: “camera movement only,” “match this performance,” or “use this clip for the timing.” Then spell out what should stay fixed and what should change. For instance: “Match the slow camera move from the reference, but use a bright studio setting and a different subject.” That makes the intended transfer clearer than simply saying “use this as inspiration.”

Put constraints at the end of the prompt. Make them observable: what stays fixed, what must not appear, and when the change happens. A Seedance prompt-template guide recommends placing constraints last and using them to specify what to keep fixed, exclude, and time. (https://wavespeed.ai/blog/posts/blog-seedance-2-0-prompt-template/)

Worked example: adapt a reference clip for a product shot

Suppose a skincare brand has a short clip of an amber serum bottle on a dark counter. It likes the slow camera push-in, but wants a pale stone setting and a cleaner product-ad look. The reference’s job is the camera movement—not the original background. A guide recommends trimming a source clip to the relevant segment before using it as a reference. (https://blog.siray.ai/seedance-2-prompt-guide/)

Brief part	Direction
Subject	One amber serum bottle
Action	A single bead of water slides down the glass; the bottle stays upright
Camera	Match the reference’s slow push-in; finish on the bottle
Setting and style	Pale stone counter, soft window light, clean product-ad look
Reference role	Follow the camera movement only; do not carry over the dark background
Constraints	Keep the bottle’s cap and label consistent; add no hands, products, or text

Copy-ready prompt: “One amber serum bottle on a pale stone counter. A single bead of water slides down the glass while the bottle stays upright. Match the slow push-in from the reference, finishing on the bottle. Soft window light, clean product-ad style. Use the reference for camera movement only; do not copy its dark background. Keep the cap and label consistent. No hands, extra products, or added text.”

Now diagnose the result. If the bottle is right but the camera is too fast, change only the camera phrase to “very slow push-in.” If the camera is right but the background stays dark, make the reference role and background change more explicit. If the bead never moves, simplify the action to that one visible event. The goal is not to keep adding adjectives; it is to make the failed instruction easier to identify and test.

If you want to practice this kind of controlled shot brief in a workflow available today, Imaginary offers Google's Veo 3.1 in Fast, Quality, and Lite modes, along with text-to-video and image-to-video. Veo 3.1 also supports first-and-last-frame control on Imaginary. Use the mode and input options that fit your test, then compare the result against the same brief. Imaginary is an independent multi-model generator, not affiliated with or endorsed by Google; the model name describes which model generates the video.

Reusable Seedance 2.0 prompt-and-reference brief

Fill this out before generating. Leave any field blank if it is not important to the shot.

Subject: [one main subject and defining details]
Action: [one main visible action; sequence it if necessary]
Camera: [shot type and one move, or “locked off”]
Setting and style: [place, lighting, and visual treatment]
Reference role: [the specific movement or performance to follow]
Keep fixed: [details that must remain consistent]
Change or exclude: [what should differ; what must not appear]
Timing: [when the action starts, changes, or ends]

The structure combines subject, action, camera, style, and constraints, while keeping the reference role separate. It is a starting point: simplify it if a field adds no useful direction. Prompting guidance also recommends keeping the number of elements manageable and using specific motion and filmmaking terms. (https://wavespeed.ai/blog/posts/blog-seedance-2-0-prompt-template/) (https://sagnikbhattacharya.com/blog/better-prompts-seedance)

Checklist

• Is there one clearly identified main subject?
• Does the prompt name a visible action rather than only a mood?
• Is the camera instruction specific, or intentionally fixed?
• Have I stated the reference’s job in a short phrase?
• Do the prompt and reference agree on movement and timing?
• Have I said what should stay fixed, change, or be excluded?
• If using a source clip, have I selected the relevant segment?
• After the first result, will I change one major instruction at a time?

FAQ

What should a Seedance 2.0 prompt describe?

Describe the subject, its main action, camera behavior, setting or style, and any constraints that protect the brief. Make timing explicit when the shot depends on a particular order of events. (https://videoweb.ai/blog/detail/Seedance-2-0-Video-Generation-Guide-Tutorial-Prompts-578fb91b8f46/) (https://blog.siray.ai/seedance-2-prompt-guide/)

What should go in a Seedance 2.0 reference instead of the prompt?

Use a reference video to anchor a movement, camera move, or performance when that is the part you need to match. Use the prompt to explain the reference’s role and direct what should happen differently. Check that your chosen tool accepts the reference type you intend to use. (https://blog.siray.ai/seedance-2-prompt-guide/)

When should I use text-to-video versus reference-to-video?

Use text-to-video when a written description is enough to invent the scene. Test reference-to-video when a particular camera move, action sequence, or performance needs to match a vision. For strict continuity, a VideoWeb guide recommends beginning with text-to-video only when continuity is not the main requirement. (https://blog.siray.ai/seedance-2-prompt-guide/) (https://videoweb.ai/blog/detail/Seedance-2-0-Video-Generation-Guide-Tutorial-Prompts-578fb91b8f46/)

How do I stop a reference from changing parts of the shot I want to replace?

State what the reference is for, then name what to keep and what to change. For example: “Follow the reference’s slow push-in only; use a pale stone background instead of the dark counter.”

How do I fix a result that misses the brief?

Classify the miss as appearance, action, or camera. Edit the instruction or reference role tied to that miss, then test again without changing unrelated parts of the prompt.

Put the brief to work

A focused shot brief makes it easier to test an idea and understand what needs revision. To apply the same prompt discipline to a video workflow you can use today, try Imaginary: it provides new users with free credits to start, and offers Google's Veo 3.1 in Fast, Quality, and Lite modes. Imaginary is independent and is not affiliated with or endorsed by Google; the model name identifies the generator. Compare your result with the brief, adjust one instruction, and keep the version that better matches your intent.

Start creating with Imaginary

Imaginary lets you generate AI video with native audio, up to 4K, from one account with credits that never expire — and free credits to start. Veo 3.1 is live now, with Seedance 2.5 and Wan 3.0 coming soon.

• Generate your first video now: Get started
• Compare plans and pay-as-you-go pricing: See pricing