AI Video Generation Model SeaDance 2.5

First, the maximum generation length of a single video is 30 seconds, ranking first in the world.

Video creation supports the joint input of up to 50 omni-modal assets.

At the same time, it has more flexible and controllable video editing capabilities.

The TV series "The Ferryman: Legends of Nanyang"

Since waking up this morning, multiple abnormal events

have been reported around the world.

You sleep in the darkness.

You are with me.

What happened?

He was isolated into a paper cutout.

Don't twist him.

"The Ferryman: Legends of Nanyang"

No, no, no, no, no.

No, no, no, no.

No, my prompt.

Hello?

Is it fixed yet?

We're about to play your video at the event.

There's been a small, unexpected issue.

I'll fix it right away.

Just give me a few more minutes.

You might be your worst enemy.

Out of control.

You might be your worst enemy.

It appears that some higher dimensional entity

is now repairing all of the anomalous phenomenon

that occurred today.

You might be your worst enemy.

Out of control.

Hurry, hurry, he sent it!

I'm on it.

This is the brand-new Doubao video generation model, SeaDance 2.5.

Currently, the new model is at the end of its closed beta phase,

and it will be officially released in early July.

The video just now showed you an edited AI world,

and it also reflects the three major upgrades of SeaDance 2.5 this time.

First, let's introduce the Doubao video generation model.

Over the past year or so,

video generation models have made rapid progress in realism, controllability, and consistency.

However, duration remains a critical factor limiting the development of AI video.

Similar models on the market only support a direct output of 15 to 20 seconds at most.

This time, SeaDance 2.5 has directly broken through this bottleneck,

generating a single video of up to 30 seconds.

This makes the camera expression much more coherent

and significantly improves creative efficiency and the quality of the final product.

In addition, multi-reference capability has always been one of creators' favorite SeaDance features.

This time, SeaDance 2.5 has also significantly upgraded this capability.

We support the joint input of up to 50 multi-modal assets, the most in the world.

Now, let's look at an example.

We input the image assets of more than 10 actors at once

to see how the model orchestrates them.

Subtitles provided by volunteers

Isn't the orchestration absolutely perfect?

In addition to understanding more assets at once,

maintaining the stability of the assets during the generation process

is equally crucial.

What we are actually demonstrating here is

a scenario mentioned in Lubo's speech earlier:

previz of white models in film production.

We input a very complex

white model of a spaceship with a complexity

approaching 100,000 polygons,

while providing a rendering material reference

for SeaDance 2.5 to generate a rendered video

to simulate the corresponding shot.

The difficulty of this task lies in the fact that

the spaceship itself contains a lot of details,

including structure, proportions,

and fine designs.

Therefore, during the camera's

slow push-in and complex movement,

we must maintain this structure and proportion,

while also completing the corresponding

generation of materials, lighting, and space,

as well as the camera movement.

From the results, we can see that

the main outline of the spaceship,

its proportional relationships, and complex structure

have all been maintained very stably.

This also means that

in professional creative workflows,

the large amount of assets, designs,

and camera movements invested by creators early on

can be stably carried over by the model,

making creation more controllable and efficient.

Finally, let's take a look at

the editing capabilities of SeaDance 2.5.

Our new model supports highly flexible

video editing functions.

Creators can keep the overall image unchanged

while modifying specific parts of the content.

For example, fine-tuning the background, replacing the demo product, and changing the model.

Let's look at an example of an ad.

Today, I'll use color 11.

Today, I'll use color 11.