Introducing Character World Model-1 the most controllable interactive avatar model in the world Features and how we built it 👇️

2/ Here is a live AI conversation with video, voice, and real-time text. The avatar is powered by CWM-1.

3/ Try CWM-1 for free! Take a selfie and make yourself dance 💃

4/ The biggest update in CWM-1 is the Emotion Engine, which controls how a character acts and moves. The engine models the internal state of the avatar, and then predicts emotions and actions for the avatar based on this.

5/ CWM-1 supports whole-body actions and hand-object interactions on any character with ZERO training.

6/ To build CWM-1, we had to solve multiple hard problems, from training our own video DiT to solving error accumulation for multi-hour long streams. We detail these and more in our research page:

Congrats on the launch! Would love to have it in ComfyUI!

Let's talk!

This model is amazing! Great work team 👏

🙌

🚀🚀🚀

🚀

Congrats team! Great work!

👏

Hell yeah @lina_colucci!! So excited to try it

💯🤯💯

emotion engine = 🤯

this is amazing. Bryce the goat designer.

100%! @designbybryce is amazing

This isnt some insanely expensive pre-recorded demo, try it on our launch page now

it's free for everyone to try!

How do I try this out??

it's free to try here:

Finally, a world model with more emotional range than my group chat 😄

shoutout chat 😂

real time video, voice, and text combined with a controllable avatar opens up some interesting possibilities for interactive AI characters.

100%

How does the action engine know what to do?

great q! outputs are conditioned on the avatar's audio, the transcription of that audio, and the user's audio (when provided)

the eye contact in the demo is unsettling in the right way

go on... 👁️👄👁️

I did not expect the hands to look that good

some might say the hands are the best part 🫶

Most controllable avatar model sounds incredible. Can’t wait to try it

Let us know what you think!

Great Innovation

The demos look incredible

Appreciate it! 🫡

Excited to see how this was built

Check out the tech report for details! linked in 🧵

🔥🚀 🔥💯🌙🔥

Wait, what actually makes this a world model?

Great q! CWM-1 has an internal representation of the world and how characters move within a scene. This allows it to simulate much more than just a talking head.

Looks like a big step forward for controllable avatars

How long did this take to build?

more time than we'd like to admit 😂 but we're excited to finally open it up to general access! our early adopters have had really good engagement with it.

The hair touching looks so natural wow

Thanks! That's just one example of the realistic physics that comes with CWM-1

awesome

🙌

what’s the API costs per minute like 👀

check out our pricing page for a complete breakdown

Open-weights?? 👀👀

For a character world model, weights are only one piece of a bigger system 😁 CWM-1 is available today via the LemonSlice API

This could be huge for virtual characters.

is huge* 😎 Character world models are the future of embodied AI.

The level of control here is what really stands out 🔥

100%! Developers can drive the avatar state directly, or use the built in Emotion Engine.

Wow this model is superb

appreciate you! we're super proud of it

The built-in physics engine is so cool.

Thank you! One of the many benefits of training a world model for embodied AI.

This could be really useful for games and virtual worlds.

Totally. We've seen lots of adoption in the entertainment space.

Does this really work with any image?

it really does!

Really cool to see avatars becoming this expressive and interactive.

The possibilities are endless ✨🌈

Would love to learn more about the custom kernels you built.