Video yükleniyor...
Video Yüklenemedi
kimodo.cpp is up on GitHub! It allows you to animate a skeleton by describing what you want it to do.
141,847 görüntüleme • 1 ay önce •via X (Twitter)
46 Yorum

nice work.... strong follow... ive been generating lots of motion with that yall can download. this is really cool to be able to do any kind of motion... going to make interesting dynamics and ai agents in games pretty interesting and life like :)

Very cool

can it do pigs

no

This looks cool. I tried using Ardy for an llm to control an avatar in real time but it was just not at a point where it’s useful yet: self clipping, weird drifting motions between animations, somewhat mushy animations here and there. Overall impressive though.

Interesting. I think with Kimodo the training distribution is limited, but what it knows how to do it does well. But that's just my initial impression.

Yeah don’t get me wrong it was really impressive, just not actually useful yet.

Interesting!! Motions looks fluid and natural... How is it behaving with GPU frequency, does it spike it or super lightweight??

Do you mean does it require a lot of compute? There are two models; text encoder and control policy. The text encoder is full precision 8b Llama which is much heavier than I'd like, but could be optimized. The control policy is much smaller, but still significant. It's fine for offline processing, realtime will be difficult on average hardware

Humanoid or also other shapes?

Practically just humanoid for now. NVIDIA haven't released a quadruped model or the training code. So it would require recreating the training code, creating a data-set for a non-humanoid etc. Completely possible, just would take me a lot of work.

super cool, but why llama, it's needs approval on hugging face, why fucking llama

This is NVIDIA's choice, but I've uploaded the GGUF to HF now so no approval needed

it works! thanks

Looking really impressive! can't wait to try it out

Nice )

you beat me to it xD

Please add an open source license to it. I would like to play around with this. Very good work!

oh nice, I gotta try this

so nice

Thanks for sharing this. Looks really cool.

Woow💯💪

I want to hook this up to Blender -> H3

this + image2threejs would go crazy

the future looks bright

might be the quickest follow in my x history

bro this is amazing

Holy fuck I've been trying to do this for months.

@toyxyz3 Wow! Dose it work with non humanoids?

@toyxyz3 Not without reverse engineering the training code

@toyxyz3 I’ll check it out. Gg

Is the generation limit a max of 10 seconds?

yes, but it is possible to stitch multiple prompts together, which I'm experimenting with.

describing motion in words is the easy demo. the hard part is the model knowing which constraints are physical and which are stylistic. 'walk sad' should keep the feet on the ground, 'walk drunk' gets to break that. curious where it draws that line.

mixamo compatible? Should nicely play with our rigged 3dgs avatars in @arrival_space_ then

I have this running on the web with Nira no device restrictions

O.O once again.. innovation at lightning speed

很酷!

does kimodo enforce foot contact, because text-driven motion without it breaks quickly in production

Does it work with non-humanoid forms?

Kimodo.cpp doesn't yet. I'm not sure if upstream have a quadruped. In theory it can with more training, but practically speaking this is a limitation

super cool!

what's the cycle time between prompt and animation playing? ... asking for a friend

how well does it handle the personalities? a whimsical walk or walk while carrying heavy load on your shoulder using both hands. something like that

so .. Unity's embedded animation AI generative system is a joke.

or better yet release all the training data for free as animation files .......
