DDIM sampling do not use x0_strength, and always iterate for ddim_steps=20, time consuming.

[Sutongtong233]

Rerender_A_Video/src/ddim_v_hacked.py

Line 300 in dfaf9d8

for i, step in enumerate(iterator):

When the x0_strength is small, it will be a long waste of time to reach

Rerender_A_Video/src/ddim_v_hacked.py

Line 306 in dfaf9d8

if strength >= 0 and i == int(

Previous blended img is useless,

Rerender_A_Video/src/ddim_v_hacked.py

Line 308 in dfaf9d8

img = self.model.q_sample(x0, ts)

Maybe we can directly:

for i, step in enumerate(time_range):
    index = total_steps - i - 1
    if strength >= 0 and i == int(total_steps * strength) and x0 is not None:
        ts = torch.full((b, ), step, device=device, dtype=torch.long)
        break
img = self.model.q_sample(x0, ts)

to get xt, and then denoise from this timestep. For controller, we always fetch the last item:

x0 = F.instance_norm(x0) * self.step_store['first_ada'][-1] + self.step_store['first_ada'][-2]

[williamyang1991]

Yes, we just use this naive implementation for quick code devolopement.
In our diffusers implementation, we have fixed this wasting iterations