Planning a 30-Second AI Short with Seedance 2.5: References, Key Frames and Draft Passes

Seedance 2.5 is ByteDance’s next-generation video model, and it gives storytellers two things that matter for narrative work: room to tell a beat in up to 30 seconds and native 1080p output.

This guide covers how to plan a short around it, from reference boards to draft previews.

What changes for storytellers

A single generation can run from 4 to 30 seconds, long enough for a setup, a turn and a payoff.
Output goes up to native 1080p, with 480p and 720p for cheaper exploration.
One request can carry up to 30 reference images, 10 reference videos and 10 audio clips.
A draft mode renders a 480p preview first, so you can check a shot before committing to the final render.
Audio generation is on by default, so a clip can arrive with synchronized sound.

Chapters

Why 30 seconds matters for story structure

Most AI video work so far has been built from five- or ten-second fragments. That length suits a reveal or a camera move, but a scene with a character who wants something, meets resistance and reacts needs more time.

Seedance 2.5 was introduced at ByteDance’s 2026 Volcano Engine FORCE conference with native 30-second generation as one of its headline features. For a storyteller, that changes the unit of planning. Instead of writing a shot list of twelve fragments, you can write three or four scenes and treat each as a single generation. Fewer seams means fewer places where costume, lighting or the character’s face drift between cuts.

That does not make every scene a 30-second scene. Duration is a parameter from 4 to 30 seconds, and you can also pass -1 to let the model pick a length that suits the task. A quiet establishing shot might want 6 seconds; a dialogue exchange might want 20.

Choose one of three starting points per scene

The model offers three ways to anchor a generation, and they cannot be mixed in a single request. Picking the right one for each scene is the most important planning decision.

First frame only

Give the model an opening image and a prompt, and it animates forward from there. This works for scenes where you care about where the shot starts but want the model to find the motion. A storyboard panel or a piece of concept art makes a good first frame.

First and last frames

Supply both an opening and a closing image. The model fills in the movement between them. The API documentation recommends this mode when the opening and closing shots must match specific images exactly. It suits transitions: a character walks from a doorway to a window, and you already know both compositions. A last frame always requires a first frame.

Multimodal references

This mode takes up to 30 images, 10 videos and 10 audio clips as references rather than as fixed frames. The model reads them for characters, props, color, framing and even movement. This is where continuity work happens. Feed the same character sheet, the same costume photos and the same location stills into every scene, and the model has a shared visual vocabulary to draw from. Reference videos and audio clips each have a combined limit of 30 seconds.

Build a reference kit before you write prompts

Because references carry so much of the continuity burden, it pays to assemble them first. A practical kit for a short might look like this:

Character sheet: front, three-quarter and profile views of each lead, in the costume they wear for most of the film.
Location stills: two or three angles of each set, lit the way the scene should be lit.
Prop close-ups: anything the plot turns on, such as a letter, a key or a phone screen.
Motion reference: a short clip showing a camera move or a gesture you want echoed.
Sound reference: a room tone or musical cue if you want the generated audio to sit in a particular mood.

Images should be JPEG, PNG or WebP under 30 MB each. Reference videos can be MP4, MOV or MKV up to 200 MB, and audio can be WAV, MP3, AAC or OGG up to 15 MB.

Use draft passes to protect the final render

Long generations are where mistakes get expensive. A 30-second 1080p clip that turns out to have the wrong blocking is a lot of render time thrown away. Seedance 2.5’s draft mode addresses this. With draft enabled, the model renders a 480p preview. If the preview works, you pass its task ID to the companion draft model (bytedance/seedance-2-5-draft) and request the 1080p version. The draft task ID stays valid for 7 days, which gives a small team time to review previews together before anyone commits.

A sensible rhythm for a short is: write the scene, attach the reference kit, generate a draft, review it with whoever owns the story, adjust the prompt or references, and only then render the final. Treat drafts the way an animator treats a rough pass.

Write prompts like a director, not a tagger

The prompt field accepts up to 30,000 characters, so there is no reason to compress a scene into a string of keywords. Describe what happens in order: who is in frame, where the camera sits, what moves first, what the character does, how the scene ends. The model is described as capturing framing and cinematic language from its inputs, and it supports direction around camera movement and performance. Phrases like “slow push-in”, “over-the-shoulder” or “she hesitates before answering” give it something concrete to act on.

A scene prompt for a 20-second beat might read: “Interior, small kitchen at night, warm practical light from a single pendant lamp. Camera starts wide from the doorway. The older woman from the reference sheet sits at the table reading the letter from the prop reference. She folds it slowly, looks toward the window, and the camera pushes in to a close-up as she exhales. Quiet room tone, a clock ticking.”

Getting access without managing infrastructure

Seedance 2.5 is available through hosted API providers rather than as downloadable weights. For writers and small studios who want to script the process, the Seedance 2.5 API page on Kie.ai has a playground for testing prompts and references in the browser, plus the full parameter list for moving those settings into code. Requests go to a single task endpoint with the model name bytedance/seedance-2-5, and the same endpoint also serves Seedance 2.0, 2.0 Fast and 2.0 Mini. That makes it simple to compare an older model against 2.5 on the same scene.

Generation is asynchronous. You submit a task, then either poll for its status or provide a callback URL that receives the result when the render finishes. Finished files are kept for 14 days, so download anything you plan to edit. Some video tasks run longer than 10 minutes, and a pipeline should keep checking until the state reads success or fail.

Practical limits to plan around

Modes do not combine. If a scene needs a locked opening frame and a set of character references, either use first-and-last frames, or use references and describe the opening composition in the prompt.
Reference video adds billable time. With reference videos attached, billing counts the reference duration plus the output duration at a lower per-second rate. Reference images and audio add no billable seconds.
Audio is on unless you turn it off. If you plan to score the short yourself, set generate_audio to false to keep the track clean.

Where this leaves the AI short

Seedance 2.5 does not write the story, and it will not fix a scene that has no conflict. What it changes is the cost of keeping a story coherent on screen. Longer generations reduce the number of seams, references carry characters from scene to scene, and draft passes let you check blocking before paying for a full render. For storytellers who have been fighting continuity one five-second clip at a time, those are practical gains.

If you want to experiment alongside other models, Kie.ai offers Seedance 2.5 together with other video, image and language models under one API key, which keeps a mixed pipeline manageable. Start with one scene, one reference kit and a draft pass, and build the workflow out from there.

Master the Art of Video Marketing

AI-Powered Tools to Ideate, Optimize, and Amplify!

Spark Creativity: Unleash the most effective video ideas, scripts, and engaging hooks with our AI Generators.
Optimize Instantly: Elevate your YouTube presence by optimizing video Titles, Descriptions, and Tags in seconds.
Amplify Your Reach: Effortlessly craft social media, email, and ad copy to maximize your video’s impact.

Start for free!