about hardware requirement and checkpoint release

[NeutrinoLiu]

Hi great project! thx for releasing the code!

I have two questions, first is there any minimum requirement for each step (data preprocess, generative model training, and model inference). Sad truth is that not every lab has a bundle of 8x80G GPU.
Another question is, instead of training our own generative model (vae, diffusion, GSM), will it be possible for you to release the well trained checkpoint directly?

Thx for your great effort again!

[yifanlu0227]

Hi!
The minimum requirements would be

1. data preprocessing: at least 24 GB VRAM GPU. You can shrink the chunk size in segformer & metric3d inference to save memory
2. generative model training: 80 GB VRAM GPU. This is a must.
3. model inference: 20 GB VRAM GPU

Unfortunately, due to the company constraint, we can not release the checkpoints. Sorry for the inconvenience.