Can more keypoints increase driving scale? long hair cannot be driven correctly

[jcn16]

[cleardusk]

More keypoints maybe better, but its hard to specify the best number of keypoints. Honestly, non-diffusion based methods are hard to simulate the hair swaying.

[molegod]

@cleardusk Is there anyway to specify keypoints in the current implementtaion?

[jcn16]

Thanks for your reply. I visulized the implicit keypoints and found all keypoints are close to face region. I wonder whether driving scale can be enlarged by adding more fluffy hair or long hair data.

[jcn16]

@cleardusk Is there anyway to specify keypoints in the current implementtaion?

I guess you can specify keypoints by "Landmark-guided implicit keypoints optimization" mentioned in paper.

[oneThousand1000]

Thanks for your reply. I visulized the implicit keypoints and found all keypoints are close to face region. I wonder whether driving scale can be enlarged by adding more fluffy hair or long hair data.

Hi, how did you visualize the implicit keypoints? Did you do some 3D-to-2D things? I am confused because there seems to be no projection matrix in the project.

=== update ===

Finally, I tried to visualize all implicit points and found that they seem to have no explicit semantic information. I also tried to change those points with 3D keypoints (estimated by 3DMM), the results were bad and unrealistic. I think the model should be re-trained to achieve explicit control.

[xll-ecnu]

Thanks for your reply. I visulized the implicit keypoints and found all keypoints are close to face region. I wonder whether driving scale can be enlarged by adding more fluffy hair or long hair data.

Hi, how did you visualize the implicit keypoints? Did you do some 3D-to-2D things? I am confused because there seems to be no projection matrix in the project.

=== update ===

Finally, I tried to visualize all implicit points and found that they seem to have no explicit semantic information. I also tried to change those points with 3D keypoints (estimated by 3DMM), the results were bad and unrealistic. I think the model should be re-trained to achieve explicit control.

can you provide me the implicit keypoints visualizing code? I wanna do the same thing! thank you!