A practical reference stack for Seedance 2.5 motion

A 30-second AI video clip asks a generation model to hold onto more than a look. It has to carry a subject through several actions, preserve a location, and make its sound fit the changing picture. A long prompt can describe all of that, but it is a poor place to manage a pile of visual and audio examples.

A more workable stack is to use ScriptFrame to break the idea into storyboard scenes, then use Seedance 2.5’s multimodal references to guide a selected render. ScriptFrame supports Seedance 2.5, and its timeline editor can preview, reorder, trim, and merge clips. Seedance 2.5 supports clips from 4 to 30 seconds and up to 50 multimodal references. That limit is capacity, not a recommended reference count.

Start with a shot that has a job

Take a simple example: a courier rides into a rain-dark tunnel, brakes beside a stalled car, and hears a warning over a radio. That is one continuous dramatic idea, but it contains separate demands: the courier’s appearance, the tunnel’s layout, a three-part motion sequence, and an audio cue that arrives at the right moment.

In ScriptFrame, turn the premise into a small number of storyboard beats before rendering. Keep the action legible: approach, brake, reaction. Specify what should remain stable and what should change. The courier’s clothing and bicycle should persist; the headlight, rain, and camera position can change with the action. A storyboard gives you a place to catch conflicting directions before they become a long render.

Give each reference a single role

Build the reference packet around jobs, not volume. For this shot, gather a small set of images for the courier and bicycle, a few for the tunnel’s shape and lighting, and selected frames that communicate the approach, stop, and reaction. Add an audio reference only if it carries a useful cue, such as the radio warning or the rhythm of the scene.

If the shot needs many examples, label your working collection by purpose before sending it to the renderer: identity, location, motion beat, or sound. That organization is for the creator; do not assume the model will follow a particular order or interpret labels as controls. Avoid near-duplicate images. They consume space without adding much new information, and contradictory references can muddy the result.

You can use up to 50 multimodal references, but begin with the few that best explain the shot. Add more only when a test reveals a specific gap. A character image can anchor appearance; it does not necessarily specify how the character moves. A frame can show a pose; it does not guarantee a precise camera path. Think of references as evidence about the intended result, not a set of hard constraints.

Separate motion from continuity

For a complex move, write the motion as a short sequence in the storyboard: the courier enters from the left, slows near the car, then turns toward the radio. Keep camera direction and screen position consistent in the description. Use reference images to clarify the look of the courier, bike, and tunnel; use the written beats to explain what happens over time.

ScriptFrame’s consistency rendering can help keep characters, props, and locations consistent across scene states. Use that for continuity planning rather than expecting one crowded reference packet to solve every state change. If the scene needs several distinct actions or a major change in viewpoint, split it into storyboard scenes and render shorter clips. Short tests are easier to diagnose than a full-length attempt.

Use audio as a cue, not a promise

Audio references can help communicate the texture or timing you want, while Seedance 2.5 supports synchronized audio in its longer text-to-video workflow. Still, a sound example is not a guarantee that a specific word, beat, or effect will land on an exact frame. State the intended event in the scene description, keep the cue simple, and inspect the rendered result. If the sound and action do not line up, revise the shot or break it into smaller beats instead of adding unrelated audio samples.

For background on the broader shift toward longer clips and synchronized sound, see the market digest on 30-second clips and synced audio.

Review, then spend references deliberately

Generate a test, then check three things separately: does the subject still look like the reference, does the action read in the intended order, and does the sound support the visible event? If identity slips, strengthen or simplify the visual references. If the motion is confused, reduce the number of beats or make their order clearer. If the audio distracts, remove samples that do not serve a specific cue.

Bring the usable clips into ScriptFrame’s timeline to preview, trim, reorder, and merge them. The trade-off is straightforward: more references can provide more context, but they also create more opportunities for mixed signals. A disciplined packet and a storyboard are usually a better starting point than filling all 50 slots.