Additional Loss term compared with paper

[CAI Zeyu]

May I ask about the first term of the loss function? (sigma_t / (sigma_t_next + 1e-9)) * (pred_epsilon - noise_pred) It seems that it always equals zero.

Consistent3D/threestudio/models/guidance/stable_diffusion_guidance.py

Line 394 in 9118a5c

 grad_latents = (sigma_t / (sigma_t_next + 1e-9)) * (pred_epsilon - noise_pred) + (noise_pred_next - pred_epsilon) 

[Zike Wu]

It is borrowed from Common Diffusion Noise Schedules and Sample Steps Are Flawed. You can optionally activate it by setting "recon_std_rescale" to non-zero integers. However, we found it not suitable in our work, so we set this rescale to zero as default.