Open bigboss97 opened 4 years ago
Taichi have very poor Generalization, so basically it only works with other Taichi videos.
Taichi have very poor Generalization, so basically it only works with other Taichi videos.
Is there any way to train it ourselves, so it would work with our input data?
Try it out.
Try it out.
Can you give us the command? I found something in the description. But I haven't really understood.
Is there a better model already made for full body motion?
Is there a better model already made for full body motion?
I've tried Impersonator++, not bad at all. https://youtu.be/91D1-Zl-w_8
This is the tutorial: https://youtu.be/Io1oEg7jIWA
There's a problem in that. The body is always in the centre. So it looks like the dancer is sliding, i.e. moonwalking 🤣 Here is how I fixed the sliding with green screen: https://youtu.be/DWKbb6LPiqw A lot of editing work. I wish someone can fix that.
I tried to follow the instruction. But I failed to get the driver video. So I tried to make my own one. This is the command I've used (probably wrong) 😄 python /bin/first-order-model-master/demo.py --config E:\bin\first-order-model-master\config\taichi-adv-256.yaml --driving_video .\driver.mp4 --source_image .\pose.png --checkpoint E:\bin\first-order-model-master\play\taichi-adv-cpk.pth.tar --adapt_scale
and this is my result hahaha https://www.reddit.com/r/MediaSynthesis/comments/hy6we5/i_think_my_firstordermodel_taichi_has_failed/?utm_source=share&utm_medium=web2x
I've moved very slowly in the video (only little motion blur in my hands) I've used a square video and a square image. The starting poses are very similar. What else do I have to do? Does the driver video need a clear background?