Seedance 2.5 Prompts: How to Prep Your Workflow Now

ByteDance started rolling Seedance 2.5 out on Jimeng AI and Doubao Pro on July 31, with API access flagged as coming soon on BytePlus ModelArk. I have sat through enough model rollouts to know exactly what happens next. The API opens, everyone pastes in the prompts that worked on the previous version, and the first week of credits goes to discovering that the prompt shape changed underneath them.

Seedance 2.5 moves two things that sit upstream of every prompt you have written so far. A single generation now runs up to 30 seconds instead of a few, and a single pass accepts a structured stack of references instead of one or two images. Both of those change what a good prompt actually looks like.

So this is a prep guide, not a review. Seedance 2.5 is not on Segmind yet. What follows is the work your team can do this week, on models you already have API access to, so that day one of Seedance 2.5 is a migration instead of a discovery exercise.

What actually changed in Seedance 2.5

Two numbers from ByteDance's own announcement matter more than everything else in the coverage.

The first is 30 seconds in a single pass, with multi-round extension on top of that for longer pieces. Under the previous generation the working unit was a shot. You wrote a shot, you got a shot, and a sequence meant either packing a shot script into one prompt or stitching clips in post. At 30 seconds the working unit becomes a scene with an arc, and an arc is a different writing job.

The second is the reference budget, and this is where most of the early coverage is quietly misleading. The headline number is 50 references. ByteDance's Seed blog breaks that figure down: up to 30 images, 10 video clips, and 10 audio clips in a single pass. That is not one pool of 50 slots you fill however you like. It is three separate budgets. Planning against the wrong one is how a team arrives on launch day with 50 stills and nowhere to put 20 of them.

There is a third change worth planning around: clay render referencing. You pass textureless 3D blockouts and the model reads spatial structure, character poses, motion paths, and camera angles off them. If your studio already builds previs, that pipeline just became a prompt input.

Prep 1: rewrite your prompts as arcs, not shot lists

This is the change that costs the most to get wrong, because a shot-list prompt will still return something at 30 seconds. It just returns 30 seconds of a model guessing at structure you never gave it.

The habit worth building now: write the narrative arc first, in four beats, and only then attach shot detail to each beat. Opening, progression, turn, resolution. I write the arc in plain sentences before I write a single camera direction, because if the arc does not hold as prose it will not hold as video either.

Here is the shape I have moved our agency-style briefs to. Note that the beats carry time ranges and the references are tagged inline rather than dumped at the end.

For a marketing agency running 20 concepts a week, the practical move is to convert your existing brief template now. Add four beat fields, add a reference manifest field, and start filling them for campaigns you are shooting this month. When the API opens you are populating a template rather than inventing one under deadline.

One thing I have found repeatedly across long-form video prompts: the turn is the beat people skip. Opening, progression, and resolution come naturally out of any brief. The turn, the moment the scene changes its mind, is what separates 30 seconds of footage from 30 seconds of story, and it is the beat you have to write deliberately.

Prep 2: build the reference library in layers

Thirty image slots sounds generous until you try to fill them with intent. The teams that will get value out of this on day one are the ones who already have a tagged, structured library rather than a folder of stills.

I organize references in four layers, which maps cleanly onto the three budgets:

• Character: each recurring character in multiple angles and lighting conditions. This is what buys you consistency across generations, and it is the layer most worth over-investing in. Budget 4 to 6 images per character.
• Product or prop: geometry and label from several angles. For anything with text on it, get a clean straight-on frame in there.
• Scene and grade: the location and the color treatment. Two to four images does most of the work.
• Rhythm: this is the video and audio budget. A clip that carries the pacing you want, and an audio bed that carries the tempo.

For a film studio, the layer that pays off fastest is the clay render one. If you are already doing previs blockouts in Blender or Maya, export the textureless passes and start filing them alongside the shot. ByteDance's own framing is that you can commit to a camera move and a composition before generation, which is exactly the control that has been missing from text-to-video for previsualization work. A director approving a blockout is a normal Tuesday. A director approving a paragraph of prompt text is not.

The audit worth running this week is dull and useful: take your last three finished projects, and try to assemble a 30 image, 10 video, 10 audio manifest for each from assets you already own. Wherever you come up short is a gap in your library, and you have a few weeks to close it before it costs you generation credits.

Prep 3: rehearse on Seedance 2.0

You cannot call Seedance 2.5 through Segmind today. You can call Seedance 2.0, and it accepts the same conceptual inputs at a smaller scale: a prompt, a reference_images array, reference_videos, reference_audios, plus first and last frame anchors. That makes it a legitimate rehearsal surface for everything above.

The current spec tops out at 15 seconds per generation, with resolution options from 480p up to 4k and aspect ratios including 21:9. So you rehearse at half length. Write the four-beat arc, compress it to 15 seconds, and see whether your reference manifest actually steers the output the way you assumed it would.

import requests

resp = requests.post(
    "https://api.segmind.com/v1/seedance-2.0",
    headers={"x-api-key": "YOUR_API_KEY"},
    json={
        "prompt": open("arc.txt").read(),
        "reference_images": [
            "https://cdn.example.com/character-a-front.jpg",
            "https://cdn.example.com/character-a-profile.jpg",
            "https://cdn.example.com/room-grade.jpg",
        ],
        "duration": 15,
        "resolution": "1080p",
        "aspect_ratio": "16:9",
        "generate_audio": True,
        "seed": 42,
    },
    timeout=300,
)
open("rehearsal.mp4", "wb").write(resp.content)

For a production house or an MCN running volume, the useful thing to measure during rehearsal is not output quality. It is how much of your reference manifest survives contact with the model. Pin the seed, change exactly one reference layer at a time, and log which layers actually move the output. Character references that do nothing at 2.0 are unlikely to suddenly do the right thing at 2.5, and you will have learned that on 15 second clips rather than 30 second ones.

That per-layer log is the single most portable artifact you can build right now. Prompts get rewritten. An understanding of which reference layers steer your particular content does not.

Where Segmind fits

Seedance 2.0 is live on Segmind today as a synchronous endpoint, so a POST returns the MP4 directly with no polling to build. There is also Seedance 2.0 Fast for cheaper iteration passes, which is where I would run the reference-layer ablation rather than burning full-quality generations on it.

On cost, the published rate card for Seedance 2.0 is token based, and the useful detail is that input tokens are free on both paths. What you pay for is output, and the rate depends on what you fed in: $7.00 per million output tokens on a text or image input generation, and $4.30 per million on a video input one. Segmind lists an average cost of roughly $1.21 per generation across typical configurations. Rehearse on short, low-resolution clips and that number drops considerably.

When Seedance 2.5 does land on Segmind it will sit behind the same api.segmind.com/v1/{slug} pattern and the same API key, so the integration work you do against 2.0 carries over largely untouched. Keep an eye on the model library for the listing.

What I would not assume yet

Everything above is built on ByteDance's announcement and on what is verifiable on Seedance 2.0 today. A few things are genuinely unknown, and I would not build a production plan on any of them.

Pricing for 2.5 has not been published. Assume 30 second 4K generations cost meaningfully more than 15 second 720p ones and budget accordingly, but do not put a number in a client quote yet. Availability outside Jimeng and Doubao is still "coming soon" on ByteDance's side, and dates like that move. The 180 second ultra-long mode has been spotted in beta but is not something I would design a deliverable around. And the reference tag syntax may well differ between ByteDance's own surfaces and third-party API providers, so treat @Image1 as the concept rather than the literal string.

FAQ

How do Seedance 2.5 prompts differ from Seedance 2.0 prompts?

Seedance 2.0 prompts describe a shot or a short shot script. Seedance 2.5 prompts describe a 30 second narrative arc, structured as opening, progression, turn, and resolution, with shot detail attached to each beat and references tagged inline.

How many reference inputs does Seedance 2.5 accept?

Up to 50 in a single pass, split across three budgets: 30 images, 10 video clips, and 10 audio clips. They are not interchangeable, so plan each budget separately.

Can I use Seedance 2.5 on Segmind right now?

No. Seedance 2.5 is rolling out on ByteDance's own surfaces first. Seedance 2.0 and Seedance 2.0 Fast are available on Segmind today and share the same input concepts, which makes them a practical rehearsal surface.

What is clay render referencing?

You pass textureless 3D blockouts as a reference, and the model reads spatial structure, character poses, motion paths, and camera angles from them. It lets you lock composition and camera movement before generating a frame.

How long can a single Seedance 2.5 generation be?

Up to 30 seconds in one pass, with multi-round extension for longer sequences. An experimental 180 second mode has been reported in beta but is not generally available.

What should my team do this week to prepare?

Three things: convert brief templates to the four-beat arc structure, audit whether you can assemble a 30/10/10 reference manifest from existing assets, and run a reference-layer ablation on Seedance 2.0 to learn which layers actually steer your content.

Getting ready

The teams that will look fast in the first month of Seedance 2.5 are not the ones with early access. They are the ones who spent the wait rewriting briefs as arcs, tagging a reference library into layers, and learning on a cheaper model which of those layers actually does anything. None of that work requires the new model, and all of it survives whatever the final API looks like.

Start the rehearsal on Seedance 2.0 today, keep the log, and be populating a template on launch day rather than writing one.