ByteDanceLive in Snaptalebytedance/seedance-2.5

Seedance 2.5 it speaks now.

ByteDance's flagship video model, and the first one in Snaptale that generates its own sound. Dialogue arrives with the mouth already moving, up to 30 reference images hold your cast together, and the camera still behaves like someone was directing it.

auto-routed · refined in chat
take 01 · ambient16:9 · 720p
  • Native audio
  • Lip-synced dialogue
  • 30 reference images
  • Character continuity
  • Multi-shot scenes
  • 9:16 native
  • 16:9 native
  • 720p · 4–15s
02 — Key capabilities

What it's actually good at, scene by scene.

01Native audio

The sound is generated with the picture, not after it

Dialogue, room tone and effects come out of the same pass as the image, so nothing drifts. Put a line in double quotes and the character's mouth is already shaped around it — no separate TTS, no manual sync.

Single render · live clip
·01
·02
·03
02Reference sets

Thirty references, one coherent scene

Seedance 2.0 took nine reference images. 2.5 takes thirty — the whole cast, the wardrobe, the product and the location, all bound to one generation. Snaptale wires them from the assets you @-mention, in the order the prompt names them.

Three frames · same subject
Reference still used as the first frame
first frame · still
clip · live render
03Image-to-video

Start from a still you already approved

Pin any image as the first frame — a GPT Image 2 still, an uploaded photo, an earlier render — and optionally an ending frame to say where the shot lands. Snaptale passes the reference automatically when the same asset appears in both prompts.

first frame · still → clip · live render
·01
·02
·03
04Scene direction

Cut between angles inside one generation

Ask for an opening wide, a medium turn and a close, and 2.5 composes the transitions in the same render — with the audio carried across the cuts. Instruction following is noticeably tighter than 2.0 on multi-subject shots.

Three frames · same subject
16:9 · landscape
9:16 · portrait
05Format-native

Composes for 9:16 and 16:9 directly

Pick the format and the model frames for it — no center-crop, no head off-screen. Snaptale renders 2.5 at 720p and delivers the same scene in either aspect when you need the matched pair.

16:9 · landscape → 9:16 · portrait
03 — How it runs

Just talk. Snaptale routes.

No model picker, no provider key, no prompt engineering ritual. Describe the shot, the agent picks Seedance 2.5 when it fits — and you stay in the conversation while it iterates.

  1. 8s, she turns to camera and says: "You came back."
    Seedance 2.5 — staging your shot…
    in-thread
    01step / 03

    Describe the scene, and the line

    Write the shot like a director, and put any spoken words in double quotes. Snaptale recognises the dialogue and queues the clip on Seedance 2.5 so the voice is generated with the picture.

  2. bytedance / seedance-2.5rendering · 9:16 · 720p · audio on
    generation · 720p · audio off
    render
    02step / 03

    Snaptale wires the references

    Every project character, product or style asset you @-mention is passed as a labelled reference — up to thirty of them — so the cast stays on-model across the whole clip.

  3. aspect: → 9:161
    duration: 5 → 8s2
    audio: model → project BGM3
    refine
    03step / 03

    Refine without re-rolling

    Change the aspect, push the runtime, mute the generated audio and drop in your own track — Snaptale re-cuts only what changed and leaves the rest of the timeline alone.

05 — Inside Snaptale

Why run it here.

Seedance 2.5 is great. Snaptale makes it directable — persistent assets, character locks, the studio and the agent all sharing one project.

  • Audio · in one pass

    Sound that was never out of sync

    Because the audio is generated with the frames, there is no sync step to get wrong. Keep the model's track, or mute it and drop your own — the video never needs re-rendering either way.

  • Cast · up to 30 refs

    The whole cast, bound at once

    Thirty reference images is enough for a full cast, the wardrobe and the location in a single generation, so continuity stops being something you fight for shot by shot.

  • Routing · automatic

    Routed for you, and away when it matters

    Snaptale runs 2.5 as the default for safe, stylised work and routes photoreal or IP-heavy scripts to a more permissive model up front — so you don't pay for a rejected render to find out.

  • Studio · in-thread

    Studio lives next to the agent

    Trim, re-cut, swap music, change aspect ratio — in the same conversation, on the clip you just generated.

06 — FAQ

Quick answers.

01What durations does Seedance 2.5 support in Snaptale?

4–15 seconds per clip at 720p. The model itself can go to 30 seconds, but Snaptale holds every video model to the same rungs so Script planning and Timeline math stay comparable — stitch clips for longer narratives and the reference set keeps them coherent.

02Does it really generate its own audio?

Yes, and that is the headline change from 2.0. Dialogue, sound effects and background music come out of the same pass as the picture. Put spoken words in double quotes in your prompt and the model shapes the mouth around them. You can turn it off and use Snaptale's own music and voice tracks instead.

03How many reference images can I use?

Up to 30 — the widest reference set of any model in Snaptale. 2.0 took 9. Snaptale passes the assets you @-mention as labelled references in prompt order.

04Can I edit or extend an existing clip with 2.5?

Not yet. Video edit and extend run on Seedance 2.0, which lets us fix the output length — and therefore the price — before the render starts. On 2.5 the provider chooses the length of an edited clip itself, at roughly four times the per-second rate, so we have kept it off that path deliberately.

05What resolution does it output?

720p. Unlike 2.0, this model has no 1080p or 4K rung on our provider. Snaptale's exporter upsamples to 1080p with a lightweight pass before delivery.

06How does Snaptale choose between Seedance 2.5 and the other models?

2.5 is the default for safe, clearly stylised or virtual work. Scripts that read as photorealistic, or that centre on named real-world IP or public figures, are routed to Happy Horse 1.1 up front — the Seedance family's content inspection tends to refuse those, and a pre-emptive switch is cheaper than a rejected render.