this is insane! really clever compositing work combining two NeRF "passes" to get the people and environment. This is gonna be the hottest wedding service in 5 years
Looks really cool. Impressive that you managed to get a consistent video by training a model on each separate frame. Great idea to use the model from the previous frame as initialization. If you are interested, there are already multiple papers out about NeRFs for dynamic scenes (starting with D-NeRF and a more recent one would be NeRFPlayer). Really cool stuff happening in this field. I don't know of any end-user tools at the moment, but most papers published their source code. The exposure problem was also addressed by giving each image an "appearance embedding" (e.g. in Nerf-W; Nerf in the wild). Of course, by color correcting yourself, you make the job a little bit easier during training.
Thanks for the comment and references. If I get more time on this I would definitely like to explore D-NeRF and similar approaches. I figure the longer I wait the easier/better/faster the dynamic stuff will get. :)
this is insane! really clever compositing work combining two NeRF "passes" to get the people and environment. This is gonna be the hottest wedding service in 5 years
Looks really cool. Impressive that you managed to get a consistent video by training a model on each separate frame. Great idea to use the model from the previous frame as initialization.
If you are interested, there are already multiple papers out about NeRFs for dynamic scenes (starting with D-NeRF and a more recent one would be NeRFPlayer). Really cool stuff happening in this field. I don't know of any end-user tools at the moment, but most papers published their source code.
The exposure problem was also addressed by giving each image an "appearance embedding" (e.g. in Nerf-W; Nerf in the wild). Of course, by color correcting yourself, you make the job a little bit easier during training.
Thanks for the comment and references. If I get more time on this I would definitely like to explore D-NeRF and similar approaches. I figure the longer I wait the easier/better/faster the dynamic stuff will get. :)
This is great work.. following!
How?😮
I thought it was iPhone 15 lol