Pavel Konovalov

There are some weird samples in dataset. This is first close mask in the dataset. It contains two closes  Images in image-mask folder often looks like combination of two...

For me results were disappointing also for this and another one vit gan implementation. I have tried to train it for 13+ hours using 4 A100 powerfull GPU on random...

When I see some paper implementation without results in Readme, 90% there is something wrong. And often you need to rewrite to fix it. I think this vq gan implementation...

I have the same issue when using port forwarding on the remote machine. When connecting locally, the scene is rendered correctly. Moving the camera and waiting do not helps.

Thanks. Didn't you measure the loss after the buckets CFR ->hole hands CFR conversion, not for the buckets? Bucketing by itself should add extra loss. It is interesting how much.

Hi. And curious what are final/current stats? 2616 hands from readme not much

@tancik I have the same issue with my scene , when training nerfacto model when training for large number of iterations. So it is not scene specific . Will try...

@tancik Looks like the reason is not in a precision. When training with mixed_precision = False I got Nan much faster and worse quality during training for some reason.

@tancik I am using the repository main branch, but not recent. Latest commit Date: 11/2/2022 10:17:45 PM 121771d126a1c90cf39013d8016e16effbfe8efc. I have an update. My scenario: I have trained for 1h+ on...

Clear. Thanks . Does it make sense to tune nerfacto for synthetic scenes or instant ngp will work thr same or better after all?