search

cvlab-columbia / zero123

Zero-1-to-3: Zero-shot One Image to 3D Object (ICCV 2023)
https://zero123.cs.columbia.edu/
MIT License
2.59k stars 188 forks source link

Training data "T" condition #93

Closed RickyYXY closed 10 months ago

RickyYXY commented 10 months ago

I find that in your evaluation code, you add T as a condition by T = T[None, None, :].repeat(n_samples, 1, 1).to(c.device) c = torch.cat([c, T], dim=-1) But in your training code, I can't find where you add the T condition. In zero123/ldm/models/diffusion/ddpm.py, function get_input in class LatentDiffusion, the variant T has never been used since T = T[:bs].to(self.device), which means the training cannot get any pose condition at all. Is this a bug? Do you forget to add T condition here? Thks

RickyYXY commented 10 months ago

I find the variant T, sorry to interupt