SDS Loss Gradient

Problem

Compute one score-distillation-sampling gradient: w(t) * (eps_hat - eps) with weight w(t) = 1 - alpha_bar_t, where eps_hat is the frozen diffusion model's predicted noise.

The signature is sds_loss_gradient(eps_hat, eps, alpha_bar_t). Return the gradient vector.

Hint

SDS turns a pretrained diffusion model into a differentiable renderer for 3D generation.

Test cases

case 1

in: [[1,2],[0,0],0.5]

expected: [0.5,1]

case 2

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

expected: [0]

case 3

in: [[2,1],[1,2],0.25]

expected: [0.75,-0.75]

case 4

in: [[-1],[1],0.75]

expected: [-0.5]

Starter codePython

def sds_loss_gradient(eps_hat, eps, alpha_bar_t):
    # Your code here
    pass

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