Seedance 2.5 — AI Video Generator
ByteDance's new video model: 30-second one-shot clips with native audio, from up to 50 reference inputs, plus video editing and extension. Live on EvoLink — generate with Seedance 2.5 today.
≈ $1.48
≈ 100.45 cr
New accounts get 10 free credits — enough for your first clips.
What can you make with Seedance 2.5?
The scenes it does best right now — every clip is a real community generation, and the full prompt behind each one lives in our prompt library.
Cinematic film
Speaker-assigned dialogue with naturally synced lips — real acting, not just pretty frames.
UGC-style vlogs
No dialogue — just grinder, pour and cup sounds on handheld DV texture.
A complete story in one take
The camera is the protagonist — 30 seconds through a ruined Tokyo, zero cuts.
Anime shorts
A fast, fun action beat — a wakeboarder soaks a summer terrace and jumps the yachts, all from a three-sentence prompt. Action scenes are where even rival communities concede Seedance leads.
What is Seedance 2.5?
Seedance 2.0 → 2.5, measured
Seedance 2.5 is the next generation of ByteDance's Seedance video model, unveiled at the FORCE conference in Beijing in June 2026 as the successor to Seedance 2.0 — the model that currently leads Artificial Analysis' image-to-video arena.
Seedance 2.5 officially launched on July 31, 2026, first in ByteDance’s own apps such as Jimeng and Doubao Pro. Provider API access opened on August 7, and the model is now live on EvoLink — including the playground on this page.
Seedance 2.5 now powers ByteDance’s consumer apps — Jimeng and Doubao in China, Dreamina and CapCut internationally.
What's new vs Seedance 2.0
- 30-second single-shot clips, up from 15 seconds — with multi-round extension (official announcement).
- Up to 30 images, 10 videos and 10 audio files as reference inputs — up from Seedance 2.0’s cap of 9 images, 3 videos and 3 audio clips (official announcement).
- Timestamp-level editing: change a generated clip at specific time points without re-rendering the whole thing (official announcement).
How to use Seedance 2.5
Three steps to your first 30-second clip.
Open the playground
The playground on this page runs Seedance 2.5 — open it and you're ready to generate.
Write or paste a prompt
Start from a sample prompt on this page or describe your own scene: subject, action, camera, style.
Generate and iterate
Review the result, tweak your prompt, and re-run. Start from a sample prompt to see what the model can do.
Seedance 2.5 prompt tips
Four habits that save credits — from the official prompt guide and what the community learned in week one.
Fix one part with timestamp editing — don't re-roll
When one element is wrong, edit that time segment instead of regenerating the whole clip. A full re-roll risks losing everything that was already right; timestamp-level editing changes only what you point at.
The clothes of the man in dark clothes in the video are
replaced by <<<image_2_3>>>, and the clothes of the man in
light clothes are replaced by <<<image_3_4>>>.
The action remains the same, without changing the
original rhythm.Test at 480p first, then go high-res
Dial in composition and style with cheap 480p drafts, then re-run the winning prompt at 720p. Most failed generations die at the idea stage, so make failing cheap — and when a delivery needs more than 720p, upscale the finished clip instead of chasing it in generation.
draft 480p · 10s · text-to-video
final 720p · 30s · 3 reference images
deliver upscale the finished clip, not the generationSeven references are usually enough
Don't feel obliged to fill all 50 slots — heavy testers rarely used more than seven at once. Every reference is another constraint the model must satisfy; save the slots for subjects that truly recur.
the strawberry flavor referencing [reference image 1].
The opening establishes visual focus with fruit, referencing
the composition of [reference video 1], the music's downbeat
cutting in. Then the biscuits line up neatly, cut to close-up,
referencing the motion and camera moves of [reference video 2].Break long stories into short shots
A 30-second one-take is the flashy demo, not always the reliable workflow. If a long scene keeps failing, split it into shorter shots and stitch — you keep control of blocking and pacing, and a failed 10-second draft costs a third of a failed 30.
0-2s, the opening title text reads: installation tutorial
2-5s, Step 1: install the water tank, see [reference image 1]
5-9s, Step 2: install the drip tray, see [reference image 2]
9-13s, Step 3: install the collection box, see [reference image 3]The full prompt playbook lives in the library
Four more writing rules — timeline beats, reference roles, separated instructions, failure constraints — plus 100+ real Seedance 2.5 prompts you can copy and run.
Key features of Seedance 2.5
Native 30-second one-shot clips, up to 50 multimodal reference inputs, timestamp-level editing, and synced audio — here's what each one does in practice.
30-second one-shot generation
Write a scene and Seedance 2.5 renders a complete 30-second clip in one native pass — no stitching, no cuts. Characters, lighting, and camera motion stay consistent from first frame to last — and the extra length doubles as editing headroom: trimming a long take is cheaper than re-rolling short ones.
Up to 50 multi-modal references
Feed up to 30 images, 10 videos, and 10 audio clips into a single generation. Lock a character's look, a location, or a soundtrack, and it carries through every frame — the sample here runs 17 reference inputs in one continuous shot.
Timestamp-level editing
Change a generated clip at specific time points without re-rendering the whole thing. Swap costumes, replace a background, or restyle a scene — Seedance 2.5 re-renders what you point at and leaves the rest intact.
Native synced audio
Audio arrives in the same pass as the video, building on the joint audio-video architecture Seedance 2.0 introduced. Footsteps, ambience, and music land in time with the action — no separate audio step.
What to know before you plan a project
- Seedance 2.5 went live on EvoLink on August 7, 2026. Using another provider? Verify the model ID and callable route first — third-party availability still varies. Track Seedance 2.5 status updates →
- Output is 480p or 720p — those are the resolutions ByteDance's own documentation lists. 4K claims appear only on platform marketing pages and contradict those docs. Need 4K? Upscale finished clips with Topaz Video Upscale →
- Several key specs, including frame rate, remain officially unannounced.
- Seedance 2.5's dedicated face-reference policy is unannounced — upload only faces you have permission to use, and we'll update guidance here when ByteDance publishes it.
Shipping video generation inside your own product? The developer page tracks Seedance 2.5 integration status and everything technical.
Seedance 2.5 API →Frequently asked questions
Everything creators ask before their first Seedance 2.5 render.
Explore the Seedance family
Last updated: