Hi @joonpark! Thanks for these questions!
What is the right rigidity value?
In general, you have to tune the rigidity value to your dataset, mesh, and training box size. In my experience, and as you’ve seen here, the default rigidity value of 20 is far too high. The video in the guide is quite old (it’s on our list to update that page!); internally, we have observed rigidity values as low as 0.007 giving good results.
I would worry less about the actual numeric value of the rigidity constant and instead focus on the types of motion you see in your particle. Motions which look “jelly-like” or otherwise not physically realistic might indicate your rigidity value is too low. Typically these unphysical motions will start to appear before training completes, so you can check on a training job with low rigidity while it’s running and kill it early if the rigidity looks too low.
What’s going on in these plots?
In general, I recommend that you don’t pay too much attention to the training plots you posted. The one piece of information that can be useful is making sure that the Val. (rigid) (green line) is higher than the other two Val. lines (red and cyan). If all three are close, it’s another sign your rigidity might be too high. But I still recommend focusing more on:
- the distribution of particles in latent space (always reduce the centering strength until particles are spread throughout [-1.5, 1.5] in all dimensions, and
- the appearance of volumes produced by 3D Flex Generate
Initialize latents from input
In general, I recommend leaving this setting off. The latent spaces in 3D Flex and 3DVA are fundamentally different, so initializing one from the other is not typically the most useful thing to do. If you’re more interested in how they differ, a recording of a workshop in which I discuss the two is available online. I also discuss the difference a bit in this forum post.
The exception to this is if you see significant motion in 3DVA and want to try to use 3D Flex Reconstruct to recover a better map. In this case, it may be beneficial to have the 3D Flex latent space closely mirror the 3DVA latent space, but in my experience it is still best to train the data “from scratch”.
I hope that’s helpful!