ByteDance has opened access to Seedance 2.0 for a limited group of users, showing a new step forward for AI video generation. The model builds on a previous version that was already described as one of the most capable tools in the category, and the new release puts more emphasis on combining different media inputs and controlling motion from references.
The result is a system that looks important for creators, editors and AI video teams watching how quickly generated clips are moving from prompt-only experiments toward more directed production workflows.
What Seedance 2.0 can take as input
Seedance 2.0 is presented as a multimodal video generation model. That means it can work with several kinds of source material at the same time instead of relying only on text prompts.
The model can handle up to four input types at once: images, videos, audio and text. Users can combine up to nine images, three videos and three audio files, with a total limit of twelve files.
That matters because video creation is rarely just a single instruction. A user may want one image to define a scene, another frame to guide a composition, a clip to define the camera movement and audio to shape the final output. Seedance 2.0 is designed around that kind of layered direction.
The generated videos run between 4 and 15 seconds long. They also automatically include sound effects or music, which makes the output closer to a complete clip than a silent motion sample.
The reference feature is the central upgrade
According to ByteDance, the standout new feature in Seedance 2.0 is its reference capability. The model can take camera work, movements and special effects from uploaded reference videos, then apply those elements to a generated result.
It can also swap out characters and extend existing clips. The source article describes video editing tasks such as replacing or adding characters as supported as well.
This points to a practical shift in AI video. Text prompts are useful, but they can be imprecise when the user needs a specific pan, chase shot, transition, physical action or visual effect. A reference video gives the system a clearer pattern to follow.
Users can issue simple text commands that connect uploaded assets to parts of a scene. One example command in the source begins: "Take @image1 as the first image of the scene. First person perspective. Take the camera movement from @Video1." The rest of the command assigns different frames to different parts of the scene.
In plain terms, the user can record a camera movement and ask the AI model to transfer that movement into a generated video, alongside other visual elements. That is a different kind of control than asking for motion in words alone.
The demos show promise, but not the full picture
The examples shown so far come directly from ByteDance. The source article notes that they were almost certainly selected from a larger batch of generated clips, so they should be understood as best-case demonstrations rather than proof of typical performance.
The demos include a fast chase scene with a man in black clothing, a crowd pursuing him, a sideways chase shot and a roadside fruit stand being knocked over. Another example shows a girl hanging up laundry, then taking another item from a bucket and shaking it out.
A more elaborate example involves a figure reaching for a Coke, reacting to footsteps, putting it back, and a western cowboy walking away with it. The prompt also describes the background fading to black, a spotlight on a can of Coke and the line "Yikou Cola - you have to try it!"
Those examples suggest a model that can follow multi-step scene descriptions and combine motion, timing, audio cues and character behavior. But the open questions are still significant.
- No one knows yet how consistently Seedance 2.0 reaches the quality shown in the demos.
- The cost has not been disclosed in the source article.
- The time needed to generate videos is also unknown.
- Consistency remains a hurdle for professional workflows.
Those limits are important. A short generated clip can look impressive in isolation, but production use depends on repeatability. Teams need to know whether a model can preserve characters, motion, timing and scene logic across many attempts, not only in selected examples.
Access is still limited
Seedance 2.0 is currently available only as a beta on the official Jimeng website at jimeng.jianying.com. Access is not described as broadly open.
There is also a compliance restriction on uploaded materials. Realistic human faces are currently blocked, according to the source article. That restriction affects how users can test character replacement, reference footage and editing tasks involving people.
For creators, this means the most interesting parts of Seedance 2.0 may still be hard to evaluate in everyday use. The model appears capable on paper and in ByteDance’s demos, but the beta status and upload limitations make it too early to judge how it performs across ordinary projects.
The AI video race is accelerating
The release arrives just days after competitor Kuaishou unveiled its Kling 3.0 model, which also supports multimodal input and output. That timing shows how quickly the AI video market is moving.
The competition is not limited to product features. According to the South China Morning Post, the launch of these powerful video models pushed share prices of Chinese media and AI companies up by as much as 20 percent.
That market reaction reflects the broader stakes. AI video tools are becoming more capable, more multimodal and more focused on controllable editing. Seedance 2.0 adds to that momentum with reference-driven generation, automatic audio elements and support for multiple media inputs in a single workflow.
The key question now is not whether the best clips look impressive. They do. The more useful question is whether Seedance 2.0 can deliver that level of control reliably, at a practical cost and within generation times that make sense for real work.