The formula is in this picture is supposed to be sampling from the q_posterior (formula in MinImagen/images/q_posterior.png)
So the normal distribution should be x(t-1) = mu(t) + sqrt(beta) * noise (beta is var here)
The code is correct in Imagen.py, which proof that the formula in the picture is wrong( should change minus to plus)
line 370: return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
The formula is in this picture is supposed to be sampling from the q_posterior (formula in MinImagen/images/q_posterior.png)
So the normal distribution should be x(t-1) = mu(t) + sqrt(beta) * noise (beta is var here)
The code is correct in Imagen.py, which proof that the formula in the picture is wrong( should change minus to plus)
line 370: return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise