🌊 Seedance 2.0
ByteDance's multimodal flagship and our top recommendation for high-quality, controllable content, including mature work. It's the first model to fuse text + image + video + audio in a single pass, with a director-grade @reference system that gives you shot-by-shot control. As of mid-2026 it is also the most capable permissive option for adult creators leaving Grok Imagine.
Watch and follow along: Generating video with Seedance 2.0 · Build a cinematic trailer in Venice Studio · Use realistic faces in Seedance. More in the Guides hub.
Why it's the Grok Imagine alternative we recommend
- Real-face references are allowed on the official platform with active consent and identity verification. This is the consent-first model Grok abandoned. (See Consent & Responsible Use.)
- Mature/suggestive content is achievable. It sits in the permissive "gray zone" rather than Grok's hard wall. Third-party API access (BytePlus/ModelArk, fal, Atlas Cloud) opens up more still.
- Unmatched control for intimate or character-driven scenes via the reference system. You decide the subject, motion, pacing, and lighting instead of rolling the dice.
Specs at a glance
Watch: Seedance on Venice
More walkthroughs on the Video Guides page.
The @reference system (the whole point)
Upload assets and they're auto-tagged @Image1, @Video1, @Audio1, etc. Then you direct the model in plain language about what to pull from each. This is the difference between "prompt and pray" and directing.
| Command | What it does | Example |
|---|---|---|
| Character ref | Use a person/character from an image | @Image1 as the main character |
| First/last frame | Set start & end frame | @Image1 as first frame, @Image2 as last |
| Motion transfer | Copy movement/camera from a clip | use the camera movement from @Video1 |
| Style transfer | Apply an image's visual style | apply the art style of @Image3 |
| Audio sync | Sync to a track / clone a voice | sync to the music in @Audio1 |
| Multi-character | Multiple distinct refs | @Image1 is A, @Image2 is B |
Video references for physical coherence
A video reference does more than transfer a look or a camera move. It also grounds the physics of the final clip. This is the most underrated trick on this model.
Seedance 2.0 loses physical coherence when you ask it for something outside its training comfort zone: unusual mechanics, niche sports, specific object interactions, fluids, collisions, or complex body movement. Pure text prompts leave the model to guess the dynamics, and it produces warping, sliding, limbs that pass through objects, or motion that ignores weight and momentum.
Feed a video reference that contains the real-world physics you want and the problem mostly disappears. R2V extraction gives the model a concrete motion scaffold to follow instead of hallucinating one from words. The reference clip does not have to match your subject or style. It only needs to demonstrate the physical behavior. Layer your appearance on top with an image reference.
The move: grab any clip that shows the correct weight, timing, and momentum (even rough phone footage or unrelated stock). Reference its motion and physics, then swap in your subject and style.
- Pick a reference where the dynamics match what you want, not the appearance.
- Name what to borrow: "motion and physics," "timing, weight, momentum." This tells the model to ground dynamics rather than copy the look.
- Short, clean clips work best. Crop to the moment that shows the behavior.
- Stack it with a character image (
@Image1) so identity stays locked while physics comes from the video.
The core prompt formula
Straight from ByteDance's official Dreamina prompt guide:
- Subject + Motion is the logical foundation: who does what.
- Environment + Aesthetics set tone: background, lighting, visual style.
- Audio adds ambient sound / dialogue for synchronized output.
No negative prompts. Seedance does not support "no X." State what you want ("clear sunny sky"), never what you don't ("no rain").
Optimal prompt patterns
I2V: faithful character animation
R2V: motion transfer + character swap
Edit & Extend
Text rendering (slogans / on-screen text)
Seedance renders text well. Formula: [Text] + [Timing] + [Position] + [Entrance style] + [Color/Font]. Use common words; avoid rare vocabulary and special symbols.
What Seedance 2.0 is best for
Common issues & fixes
Distilled from ByteDance's internal R2V troubleshooting guide and field testing:
| Problem | Fix |
|---|---|
| ID drift (face changes mid-clip) | Use a clean single-view reference; avoid three-view / multi-view sheets as a character ref. Add beats to the prompt. |
| "Twin" artifacts / duplicate subject | Don't feed multi-pose collages; one clear subject per character reference. |
| Unwanted subtitles / logos / watermarks | Use a clean reference; describe a plain background. |
| Periodic flicker / quality dips | Raise reference resolution; generate in 2–3s chunks for complex motion. |
| Extension quality degrades | Re-anchor with a fresh high-res frame at the junction; keep style/pacing language identical. |
| Blurry output | Input quality caps output. Use 2K/4K source images; upscale low-res refs first. |
| >4 reference characters | Known weak spot. Split into multiple generations and stitch. |
The 80/20 rule: spend most of your prep time on reference quality. The biggest single-clip quality gains come from better inputs, not longer prompts. Structure complex scenes in 2–3 second chunks.
What actually gets blocked (it's mostly faces, not nudity)
Seedance is more capable for mature content than its reputation suggests. The wall most people hit isn't an NSFW filter, it's face-detection and intellectual-property moderation:
- Face detection is the big one. A lot of "content moderated" results are the system reacting to detected real faces, not to nudity or suggestiveness. Inside Venice Studio there's a supported way to work around this, which is what unlocks much more of Seedance's range. (See the "Seedance is Blocking Realistic Faces" walkthrough above.)
- Celebrity / IP detection is enforced for compliance and isn't going away. Recognizable public figures and protected IP get blocked by design.
- The catch for API users: the face-detection workaround lives in the app, not the API. Through the raw API you'll run into face moderation far more often, which makes Seedance a weaker option for pure API workflows today.
Content policy & access reality
- Official Dreamina app: blocks fully explicit acts, sexualized minors (zero tolerance, always), non-consensual scenarios, and named real public figures. Suggestive/sensual content generally passes.
- Real faces: permitted as references only with identity verification / prior legal authorization. This is the consent gate. Respect it.
- API / third-party (BytePlus, fal, Atlas Cloud): more permissive on faces and mature content, but you remain responsible for rights, consent, and local law.