Seedance 2.5 arrived with three upgrades, and they are not cosmetic — each one changes how you actually work with the model. This episode tests all three on real tasks instead of reading the release notes aloud.

1. Multi-reference plus white-model video

Several reference images and a grey/white blockout drive the shot together. The blockout holds composition and camera; the references carry identity and material.

This matters because it splits the two jobs that used to fight each other. Previously, if you wanted a specific camera move and a specific character, the reference image had to encode both — and the model would average them, softening whichever one lost. With a separate blockout it stops guessing about framing.

Practically: you block the shot in grey, and the reference only has to answer "who" and "what material".

2. Thirty-second long-form storytelling

A single prompt now carries a narrative across roughly half a minute while keeping the subject and the scene consistent.

Half a minute is the interesting number. It is long enough that identity drift and scene drift become real problems, so the question is not "can it generate 30 seconds" but "does the subject still exist at second 28". In the episode I check exactly that: same person across the whole clip, and whether the background stays put when the camera moves.

3. Precise video editing

Targeted edits inside an existing clip, instead of regenerating the whole shot.

This is the one that changes the workflow the most. Regenerating a whole shot means re-rolling everything — the performance, the timing, the background. Editing inside the clip means you keep the takes that already work and only touch the part that does not.

What the episode actually does

Rather than demonstrating each feature in isolation, the episode runs them in sequence on the same material: block and reference the shot, extend it into a long-form beat, then go back in and fix one thing.

That sequence is where the features compound — the long-form clip gives you something worth editing, and the edit saves you from re-rolling it.

Five language editions

The episode released in five languages on the AI-KSK channel:

Chinese original on Bilibili: https://www.bilibili.com/video/BV1m7uV66EjE/

Honest notes

  • This episode has no downloadable ComfyUI workflows — it is a model-capability test, not a workflow build. Nothing to install.
  • The narration in each language edition is AI-assisted translation plus AI voice synthesis, not a human studio recording. That is stated in every video description too.
  • Not affiliated with or endorsed by ByteDance. Model licences remain with their authors.