search

SUDO-AI-3D / zero123plus

Code repository for Zero123++: a Single Image to Consistent Multi-view Diffusion Base Model.
Apache License 2.0
1.76k stars 123 forks source link

About computation resources #89

Closed qwer1793 closed 3 months ago

qwer1793 commented 4 months ago

I notice that the paper didn't mention the computation resources, I'm quite interested in training a Depth ControlNet for Zero123++ on myself with only have 4 RTX3090s. Is it enough to train?

eliphatfs commented 4 months ago

You can do a ControlNet on a single 2080Ti if you want. ControlNets usually converge with a batch size as small as 4, and sudden converges in about 10k-20k steps. This would be about 2 days on an 2080Ti :) To get the best results you need a batch size of 128-256 and train for 30k steps. In that case we would recommend 8xA100 or more setup.

eliphatfs commented 4 months ago

With small batch size the model would also be usable in general, but may be slightly less clever towards some of the control images.