search

HumanAIGC / AnimateAnyone

Animate Anyone: Consistent and Controllable Image-to-Video Synthesis for Character Animation
Apache License 2.0
14.15k stars 942 forks source link

Driving pose skeleton to approximate the length of the character’s skeleton in the reference image #62

Open learnuser1 opened 5 months ago

learnuser1 commented 5 months ago

During inference, we rescale the length of the driving pose skeleton to approximate the length of the character’s skeleton in the reference image. Does anyone know how to do this step?

mymusise commented 5 months ago

We tried rescaling through the external rectangle of the kps common to ref and motion, and tested in Moore-AnimateAnyone, but the effect was not significantly improved.

learnuser1 commented 5 months ago

We tried rescaling through the external rectangle of the kps common to ref and motion, and tested in Moore-AnimateAnyone, but the effect was not significantly improved.

I tried scaling the skeleton and it worked, but it wasn't good enough.

MoonEese commented 5 months ago

In my work, skeleton alignment was effective for quality (Ref : woman, Pose : man)

wangxr1999 commented 4 months ago

I tried to extract several key points from the dwpose and perform affine transformation to obtain a matrix. Then, I processed each frame of the pose video and achieved better results

wxhqw commented 4 months ago

I tried to extract several key points from the dwpose and perform affine transformation to obtain a matrix. Then, I processed each frame of the pose video and achieved better results

It's so greatly, i have some doubts, how to select the key points?

eightmusic commented 4 months ago

In my work, skeleton alignment was effective for quality (Ref : woman, Pose : man)

How did you achieve it? Can you share the code.

wweevv-johndpope commented 4 months ago

@eightmusic - there's this code base - https://github.com/MooreThreads/Moore-AnimateAnyone/

EnochX commented 4 months ago

I tried to extract several key points from the dwpose and perform affine transformation to obtain a matrix. Then, I processed each frame of the pose video and achieved better results

That's so great. Really want to know how to make it! Can you share the code?