Questions Regarding the style_fidelity Parameter

[ShowLo]

In stable_diffusion_reference.py, there is the following code:

var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)
std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5
mean_acc = sum(self.mean_bank[i]) / float(len(self.mean_bank[i]))
var_acc = sum(self.var_bank[i]) / float(len(self.var_bank[i]))
std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5
hidden_states_uc = (((hidden_states - mean) / std) * std_acc) + mean_acc
hidden_states_c = hidden_states_uc.clone()
if do_classifier_free_guidance and style_fidelity > 0:
    hidden_states_c[uc_mask] = hidden_states[uc_mask]
hidden_states = style_fidelity * hidden_states_c + (1.0 - style_fidelity) * hidden_states_uc

And there is a comment:

style fidelity of ref_uncond_xt. If style_fidelity=1.0, control more important,
elif style_fidelity=0.0, prompt more important, else balanced.

However, when style_fidelity=1.0, hidden_states is set to hidden_states_c. Because of the presence of uc_mask, half of hidden_states_c is unrelated to the reference image. In this case, isn’t the prompt more important? Conversely, when style_fidelity=0.0, hidden_states is set to hidden_states_uc, where both the first and second halves of hidden_states_uc are influenced by the reference image. In this scenario, isn’t the control from the reference image more important?