Seedance 2.5 References: How to Use All 30 Images, 10 Videos & 10 Audio Clips

Seedance 2.5 quadrupled the reference budget, from 12 inputs to 50. Most of the advice that followed treats that as a single number to manage. It is really three separate tools that happen to share a counter, and they are good at different things.

This covers all three: what images, video and audio references each actually do well, the numbering behaviour that silently breaks working prompts, and why the right answer is almost always around nine references rather than forty.

The budget, precisely

• 30 images
• 10 video clips, roughly 1.8 to 30.2 seconds each, capped around 30.2 seconds combined
• 10 audio clips, same limits
• Tagged in the prompt as @Image1, @Video1, @Audio1

Two behaviours govern everything else.

Numbering follows upload order, not type

This is the one that bites. Everyone assumes images and videos have separate counters. They do not: references are numbered in the order you upload them, across all types.

So reordering your uploads silently renumbers every tag in your prompt, and nothing errors. A prompt that worked yesterday now points @Image2 at the wrong picture, and the output looks like a prompt-adherence failure rather than a bookkeeping one.

Fix the upload order first. Write the tags second.

Order is priority

The model weights earlier references more heavily. Slot one should hold whatever you least want to lose. For anything involving a real person, that is the face, every time.

Spend it in layers

Think of the stack as five layers, and fill only the ones your shot needs.

A bigger production genuinely needs more, but it needs more distinct jobs, not more copies of the same job:

Images: identity, place, look

Images do the heaviest lifting, and they cover four distinct jobs that are worth keeping separate in your head.

Identity. Three angles beat six near-identical selfies. The model needs something to draw on when the camera swings around a subject; redundancy just gets averaged. Full treatment in keeping a real person's face consistent.

Environment. A location plate is dramatically more reliable than a description for anything specific: a particular stadium, a set, a room with a layout that matters. Prose is fine for generic places and poor for particular ones.

Style. One frame that defines grade and mood. One. Two style references with different colour temperatures fight, and the model resolves the fight by splitting the difference into something neither of you wanted.

Props and wardrobe. Worth their own slots when the object has to be right, because a described jacket becomes a generic jacket.

The rule that ties all four together: label narrowly. Not "@Image1 is a reference" but "@Image1 defines the woman's face and hairline only." Scoping stops the model also importing that photo's background, lighting and wardrobe as a bonus.

Video: motion, and only motion

The most misunderstood modality. A video reference is the right way to transfer camera movement, choreography and edit rhythm. It is the wrong way to transfer a subject, a location or wardrobe.

Left unscoped, the model copies everything it sees. Ask it to follow the movement in a clip of a skateboarder at a beach and you can easily get the skateboarder, the board, the clothes and the beach along with the movement.

So video references need two statements, not one:

@Video1 controls camera movement and timing only: the handheld push-in and the pause before the turn.
Do not copy the subject, the clothing, the location or the lighting from @Video1.

Practical notes:

• Short clips beat long ones. Five to ten seconds of the exact movement you want is more usable than thirty seconds containing it somewhere.
• Trim to the move. A clip that also contains three other camera moves teaches the model three other camera moves.
• Watch the cost. On some providers, reference video adds to billable input duration while images and audio do not. A long motion reference can quietly cost more than the shot it is guiding.

Audio: rhythm, voice, and the thing you actually hear

Audio is co-generated with the picture in the same pass, which is why prompt language about sound genuinely changes the edit rather than just labelling it.

Music references are for structure. Attaching a track does nothing on its own; you have to say what syncs to it. "Cuts land on the beat, final pose on the last hit" turns a music reference into an editing instruction.

Voice references set delivery and timbre for dialogue. Worth a slot when the read matters.

Described sound needs no slot at all, and this is the part people skip. generate_audio defaults to on, and a prompt that says nothing about sound gets generic ambience. Describing sound where it happens costs nothing and produces something that sounds designed:

Diegetic sound: rain drumming on the car roof, wipers squeaking on each pass, muffled jazz from the radio.
She says, almost laughing: you're late.
Traffic hum stays distant throughout.

Keep the sound description next to the event that makes it. Sound gathered into a block at the end of the prompt drifts away from its trigger.

Why nine beats forty

Every reference is an instruction. Instructions that overlap get averaged, and averaging is exactly where identity drift comes from: six photos of a face at slightly different weights and hairstyles do not give the model a better idea of the person, they give it a blend of six people.

The 50-slot ceiling exists for productions carrying many characters, products and locations at once. It is not a completeness score. If you cannot say in one sentence what a reference is protecting, it is costing you.

Common mistakes

• Reordering uploads without re-checking tags. Silent, and it breaks a prompt that worked.
• Putting the style plate first. Slot one is priority. Spend it on the face.
• Two style references. They fight. Pick one.
• Video references without a do-not-copy line. You asked for a camera move and got someone else's outfit.
• Attaching music without saying what syncs. A track with no sync instruction is decoration.
• Contradicting a reference in prose. If @Image2 is a navy suit, do not write black tuxedo. The model splits the difference and loses both.
• Filling slots because they are there. Forty references is not thoroughness, it is forty competing instructions.

Where to go next

• The prompt handbook for the grammar around all of this
• Character sheet or storyboard grid for building the reference image rather than just attaching one
• Keeping a real person's face consistent for the identity layer specifically
• Writing 30 seconds for spending the duration the references are serving

Every generation on Starrd assembles a reference stack automatically, which is how we learned most of the above, usually by getting it wrong first. If you want somewhere to run your own stack, Kie.ai exposes Seedance through a single API with per-second pricing, which makes cheap drafts practical.