Seedance 2 and 2.5: A Filmmaker’s Guide to Modes, Prompts and Limits
How to use Seedance 2 and Seedance 2.5 for film work: what changed between versions, the four generation modes, prompt structures that work, pricing, and the face filter.
Seedance is ByteDance’s video model family, and it is one of the most searched AI video models of 2026. Seedance 2.0 arrived in February 2026 with native audio and multi-shot takes. Seedance 2.5 followed on July 31, 2026 and doubled the length of a single take to 30 seconds. This guide explains what each version does, how to prompt it, what it costs, and where its limits are, from the point of view of someone making a film rather than a single clip.
Seedance 2.0 vs Seedance 2.5
Both versions generate video and sound together, including dialogue with lip sync, and both accept reference images, videos and audio. The main differences:
| Seedance 2.0 | Seedance 2.5 | |
|---|---|---|
| Released | February 2026 | July 31, 2026 |
| Take length | 4 to 15 seconds | 4 to 30 seconds |
| Native audio | Yes | Yes |
| Modes | Text, image, first/last frame, reference | Text, image, first/last frame, reference |
| References (model limit) | 12: up to 9 images, 3 videos, 3 audio | 50: up to 30 images, 10 videos, 10 audio |
| Resolution in Frameliq | 480p, 720p | 480p, 720p, 1080p |
The release date and the 2.5 reference limits come from ByteDance’s Seedance 2.5 announcement. Frameliq’s reference mode for Seedance 2.5 currently accepts up to 9 images, 3 videos and 3 audio clips, 12 in total.
The four modes, and when to use each
For a film, image to video is the default. You have already decided what the shot looks like in the storyboard, so the video model only needs to move it. Reference mode is the one to reach for when the same character must appear in a shot that has no storyboard frame yet.
How to prompt Seedance 2
ByteDance has not published an official prompt guide. The most detailed public guidance is fal’s Seedance 2.5 workflow guide, and it matches what we see in practice. Four prompt shapes work:
- A cinematic one-liner in this order: camera, subject, action, environment, look, sound.
- Field blocks with labelled sections for style, scene and action.
- Staged beats, where each beat names the frame it ends on.
- A timed shot list, such as 0 to 6 seconds, then 6 to 12 seconds.
Three rules matter more than the shape you pick:
- Put dialogue in double quotes. Quoted lines are spoken and lip synced. Say who speaks and how.
- Name concrete sounds. “Rain on corrugated steel” gives the model something to generate. “Atmospheric audio” does not.
- Refer to references by position. In reference mode, the first image is
@Image1, the first video is@Video1and the first audio clip is@Audio1.
A one-liner that follows the order above:
Slow push-in, a lighthouse keeper in a yellow oilskin turns from the lamp toward the window, storm-lit lantern room at night, cold blue light with warm lamp glow, 35mm film look. Wind against glass, a radio crackles. He says quietly, “That’s not one of ours.”
And a timed shot list for a longer Seedance 2.5 take:
0–8s: wide shot, the keeper climbs the spiral stairs, lantern swinging. 8–18s: medium shot at the radio desk, he adjusts the dial, static resolves into a tone. 18–30s: close-up on his face as the tone repeats, he whispers, “Three short. Three long.” Ends on his eyes lifting to the window.
A 30-second take is useful for dialogue scenes and oners. It is also expensive to retry. For most shots, generate 5 to 10 seconds and cut. Your edit will be tighter and each retry costs a third as much.
Limits you should know before you plan a film around it
The face filter
Seedance rejects many requests involving photoreal human faces. This follows a dispute in February 2026, when Disney and other studios sent ByteDance legal demands over Seedance 2.0 output and the Motion Picture Association followed with its own letter (Deadline’s report). In practice, stylised and animated characters hold well across Seedance shots, and photoreal lookalikes are often refused. Frameliq tells you when a take was rejected for this reason, so you can switch that shot to another model.
Where you can use it
ByteDance offers Seedance in its own apps, Jimeng and Doubao in China and Dreamina elsewhere, and availability and pricing vary by country. Developers and tools reach it through API providers such as fal. Frameliq uses fal, so Seedance 2.0 and 2.5 are available in Composer wherever Frameliq is.
What Seedance costs
On fal, Seedance 2.5 is billed by output size. At the time of writing that is about $0.22 per second at 480p and about $0.47 per second at 720p, with sound included. A 10-second 720p take costs about $4.70, so it pays to settle the shot in a storyboard frame first. In Frameliq you can pay with included Media credits or with your own fal key. Check the estimate shown on the Generate button before each take, because provider prices change.
Seedance next to the other leading models
No single model wins every shot. On the Artificial Analysis text-to-video leaderboard in August 2026, Wan 3.0 ranked first and Seedance 2.5 third, with MiniMax H3 close behind. Rankings move every few weeks, so treat them as a rough guide. A practical split:
- Seedance 2.5: long takes, natural acting, multi-beat shots with dialogue, stylised characters.
- Seedance 2.0: the same strengths in takes up to 15 seconds, at a lower price.
- Veo 3.1: short, controlled shots that follow direction closely.
- Kling 3.0 Pro and Kling O3: photoreal characters and steady faces.
- Wan 3: long takes with many reference images.
- H3 Max: fast, cheaper drafts with sound, plus lip sync from a still.
Our model comparison goes deeper on choosing a model per shot.
Using Seedance inside a film workflow
A model generates clips. A film needs the same characters, places and voices across dozens of clips. In Frameliq the story, cast portraits and storyboard come first, and each storyboard frame becomes the starting image for its shot. In Composer you pick Seedance 2.5 or 2.0 per shot, choose the mode, and the prompt is already written for that model. If you write with an AI assistant, the Claude and ChatGPT connection can rewrite every shot prompt for Seedance in one pass.
Create a Frameliq account to try Seedance next to the other models on your own script, or follow the short film tutorial first.
FAQ
What is the difference between Seedance 2 and Seedance 2.5?
Seedance 2.5 generates takes up to 30 seconds instead of 15 and accepts more references. Both produce synced sound and dialogue and share the same four modes.
Can Seedance 2 generate dialogue?
Yes. Put the line in double quotes in the prompt and say who speaks. The model generates the voice and matches the lips. For a voice that stays the same across a whole film, record dialogue separately and attach it, which Frameliq’s voice masters are for.
Why did Seedance reject my prompt or image?
The most common cause is a photoreal human face in the prompt or a reference image. Try a stylised look, or generate that shot with a different model.
How long can a Seedance video be?
Up to 15 seconds per take with Seedance 2.0 and up to 30 seconds with Seedance 2.5. Longer scenes are built by cutting several takes together on a timeline.
Is Seedance 2 good for filmmaking?
Yes, for the right shots. It is strong on acting, camera movement it directs itself, and scenes with dialogue. It is a poor fit for photoreal lookalikes because of its face filter, so most films pair it with a second model.