Question about algorithm

EnVision-Research / LucidDreamer

Official implementation of "LucidDreamer: Towards High-Fidelity Text-to-3D Generation via Interval Score Matching"

MIT License

749 stars 32 forks source link

Question about algorithm #31

Closed tuning12 closed 7 months ago

tuning12 commented 7 months ago

Thank you for your great work! In algorithm 1, the noise added is fixed (determined by unet and x_0). However, in the "train_step_perpneg" function, a random noise is added which is different from the algorithm 1.

YixunLiang commented 7 months ago

Hi, thanks for your attention.

I just checked our code, and it seems like we don't have this bug, I think there is a misleading in :
prev_noisy_lat = self.scheduler.add_noise(latents, noise, self.timesteps[ind_prev_t]) However, the ind_prev_t will always be set to 0 at each step. Apologize for it.

tuning12 commented 7 months ago

Thank you for your patient reply!

tuning12 commented 6 months ago

Hi, how do you sample the viewpoint in training? I log the azimuths in the training script and find that most of them are lower than 100. Are there any tricks on it?