Seedance 2 is now live on film.fun.
ByteDance's new video model is wired into Splice today as seedance2Fast via the Replicate path, available across the existing image-to-video and text-to-video surfaces, plus the agent Shot Sheet → Seedance 2 loop for fully-composed scene generations.
What's new
Seedance 2 isn't an incremental bump on 1.5. The architecture shifts from "generate video, dub audio on top" to a unified audiovisual model. The model produces synchronized stereo audio — background ambience, foley, score, voice — as part of the same generation, not as a second pass.
Three other shifts worth knowing about:
- Up to 9 reference images can be passed in for character continuity, location identity, prop fingerprinting. The model honors the references rather than approximating them.
- Up to 3 reference videos as motion priors — pass a clip showing the camera move or character action you want, and the new generation inherits the motion grammar.
- Up to 3 reference audios for voice cloning, sound design, or musical tone — the model conditions its native audio output on the references.
- Aspect ratios are now adaptive — the model handles 16:9, 9:16, 1:1, 4:5, 21:9 without quality fall-off, so multi-platform delivery is one generation, not three.
How to prompt it
The Seedance 1.0 prompting pattern still works — describe the scene, name the transitions, keep the camera intent simple. With Seedance 2 you can stretch further into multi-shot continuity because the references handle identity, leaving the prompt to handle action.
Scene 1: [setting + subject + action] + [camera move]
Transition: [cut/dissolve/match]
Scene 2: [continuation + new framing]
Transition: [transition style]
Scene 3: [resolution + camera return]
Audio: [implicit if generate_audio is on — the model picks ambience and pacing]
Visual style: [cinematic / anime / documentary / commercial]
A single-shot prompt is fine for atmospheric moments — let the model do the slow push-in for you:
Prompt: "Slow cinematic push-in on a vintage film camera sitting on a wooden tripod inside a quiet studio. Warm tungsten key light from camera-left catches dust motes drifting across the lens. Deep navy background, shallow depth of field, restrained handheld movement. 24fps photographic look."
A multi-shot prompt earns its keep when you want scene continuity. Seedance 2 reads each labelled scene as a beat and holds the world together across the cuts:
Prompt: "Scene 1: Wide cinematic shot of an outdoor film set at golden hour... Transition: cut on movement. Scene 2: Close-up of the camera dolly wheels rolling on track... Transition: pull back. Scene 3: High-angle wide of the entire empty film set at sundown."
Both clips above were generated through seedance2Fast on Splice with native audio on by default. Headphones in, you'll hear the ambience the model wrote.
In Splice
Pick Seedance 2 Fast in the video generator's model dropdown, or call it from the agentic API. Both surfaces use the same model id under the hood, so prompts and reference uploads compose the same way whether you're iterating in the UI or scripting through an agent.
What this changes for the workflow
The Seedance 1 article said scene labelling is everything. That's still true. What changes with Seedance 2 is that you can stop fighting identity drift — the references hold characters, locations, and motion across cuts, and the model writes its own audio bed under all of it.
For a creator working a shot list, that's the difference between three generations per scene (one for the look, one for the motion, one for the audio) and one. The compute saved goes into iteration.
Sources
- ByteDance Seed: Seedance 1.0 overview (parent model family)
- Seedance 1.0 prompting guide (the foundational patterns that still apply)
Try it
Open Splice, pick a production, choose Seedance 2 in the video generator, and write your first multi-shot prompt. If you've built a Universe with character sheets and location sheets, point a Shot Sheet at the new model and watch a sequence come together.
This is the model the agentic loop has been waiting for.
