DDIM Step

Problem

Perform one deterministic DDIM reverse step: first estimate pred_x0 = (x_t - sqrt(1 - alpha_bar_t) * eps_pred) / sqrt(alpha_bar_t), then return x_prev = sqrt(alpha_bar_prev) * pred_x0 + sqrt(1 - alpha_bar_prev) * eps_pred.

The signature is ddim_step(x_t, eps_pred, alpha_bar_t, alpha_bar_prev). All values are scalars.

Hint

DDIM removes the random noise term, giving deterministic sampling with fewer steps.

Test cases

case 1

in: [1,0,0.5,0.25]

expected: 0.7071067812

case 2

in: [0,1,0.25,0]

expected: 1

case 3

in: [2,0.5,0.25,0.25]

expected: 2

case 4

in: [1,1,1,1]

expected: 1

Starter codePython

import math
def ddim_step(x_t, eps_pred, alpha_bar_t, alpha_bar_prev):
    # Your code here
    pass

4 test cases · run your code against them in the editor