Training details for inpaint

[etiennedub]

Thank you for your nice work and new contribution!

I don't know if you plan to release a new version of your paper but meanwhile I would have some question about the training procedure and details for the inpaint model.

My main question is about the "random optical flow occlusion masks". Could we have more details about it? It is a mask where the optical flow is higher than a threshold between 2 video frames?

Also, all the training details possible would be appreciate mostly number of training steps, batch size and datasets used?

[alelordelo]

@etiennedub , where you able to train a controlnet successful?

[VladAndronik]

@etiennedub it might be the regions of occlusions, i.e. when the head is shifted, we need to inpaint what is behind etc. It is usually computed by taking two optical flows from i to i+1 and back and taking the intersection.

[lllyasviel]

yes it is occ masks

[huangjun12]

@lllyasviel dear author, how about the number of training steps, batch size and datasets used for controlnet inpaint?