Closed Yuxuan-W closed 7 months ago
Hi Yuxuan,
Thanks for your question, and in fact we use the consistent noise in the following code:
pred_x0 = x_0 + ((x_t - x_s) - beta_prod_t ** (0.5) * (e_t - e_s)) / alpha_prod_t ** (0.5)
The reason why we’re not using the consistent noise here is we want to avoid floating error caused by it so we make a simplification to not use the consistent noise. Here’s how it work:
And I’m not sure whether consistency model scheduler can use eta as well.
Understand. Thanks for this clear clarifying!
Here the eta is set to 1 in your code, but this will lead dir_xt to be always 0. Besides, I'm a bit of confused of the computation of pred_x0, it seems add a target branch latent to original image's latent and then subtract the source branch latent.
Would be appreciate for your reply!