why p_mean_variance use noise_level instead of sample_gammas like in training for time conditon of denoise function.

[zwmCool]

training:

`
def forward(self, y_0,group,y_cond=None, mask=None, noise=None):

    # sampling from p(gammas)

    b, *_ = y_0.shape
    t = torch.randint(1, self.num_timesteps, (b,), device=y_0.device).long()
    
    gamma_t1 = extract(self.gammas, t-1, x_shape=(1, 1))
    sqrt_gamma_t2 = extract(self.gammas, t, x_shape=(1, 1))
    sample_gammas = (sqrt_gamma_t2-gamma_t1) * torch.rand((b, 1), device=y_0.device) + gamma_t1
    sample_gammas = sample_gammas.view(b, -1)
    ....
    noise_hat = self.denoise_fn(torch.cat([y_cond, y_noisy*mask+(1.-mask)*y_0], dim=1), sample_gammas)

`

inference:

`

def p_mean_variance(self, y_t, t,group, clip_denoised: bool, y_cond=None)

    noise_level = extract(self.gammas, t, x_shape=(1, 1)).to(y_t.device)

    y_0_hat = self.predict_start_from_noise(y_t, t=t, noise=self.denoise_fn(torch.cat([y_cond, y_t], dim=1), noise_level))

`

[xmhGit]

I have the same question. Did you find the reason?

[kirin-from-16]

#50 i think the author explain this in here.

[xmhGit]

#50 i think the author explain this in here.

Got it. Thanks~