Classifier-Free Guidance Combine

Problem

Combine conditional and unconditional noise predictions for classifier-free guidance: eps = eps_uncond + scale * (eps_cond - eps_uncond).

The signature is cfg_combine(eps_uncond, eps_cond, scale).

Hint

Scale 1 recovers the conditional model; larger scales sharpen prompt adherence.

case 1

in: [[0,0],[1,2],3]

expected: [3,6]

case 2

in: [[1,1],[1,1],5]

expected: [1,1]

case 3

in: [[0.5,-0.5],[1,0.5],2]

expected: [1.5,1.5]

case 4

in: [[0],[1],0]

expected: [0]

Starter codePython

def cfg_combine(eps_uncond, eps_cond, scale):
    # Your code here
    pass

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