PLMS sampling is broken

[xiankgx]

Using the 768 v-diffusion model, using prompt "fruit basket".

With DDIM sampling:

With PLMS sampling:

[zwishenzug]

Can confirm.

I managed to get a better output by adding in some code from the DDIM version. There's new parts for 'v' prediction like this:

    if self.model.parameterization == "v":
        e_t = self.model.predict_eps_from_z_and_v(x, t, model_output)

    if self.model.parameterization != "v":
        pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
    else:
        pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output)

The equivalent isn't in the PLMS code yet.

But I couldn't quite get it fully correct.