search

NVlabs / CTG

Other
67 stars 9 forks source link

Some questions about diffusion model #5

Open Jeff09 opened 7 months ago

Jeff09 commented 7 months ago

Thanks for such awesome work.

I have some questions about the code and need you to help me.

  1. When setting algo_config.predict_epsilon=True, the model is output noise directly. However, when computing the reconstruct loss self.loss_fn(x_recon_selected, x_start_selected) in the diffuser.py, the ground truth is x_start_selected, which is the ground truth trajectory, not the initial noise? What is the reason behind that?

  2. In the function predict_start_from_noise, what's the formula that can compute the clean trajectory directly from the predicted noise? And also sqrt_recipm1_alphas_cumprod will get nans in the formula sqrt_recipm1_alphas_cumprod = torch.sqrt(1. / (alphas_cumprod - 1.)), alphas_cumprod is always less than 1 because it's the comproduct of numbers less than 1. Do you have the problem?

  3. Why can the diffusion model predict the clean trajectory when setting algo_config.predict_epsilon=False?

Thanks for your response.