Null-Text Inversion Step

Problem

Apply a sequence of gradient descent updates to a null-text embedding: for each gradient vector in grads, update embedding = embedding - lr * grad in order.

The signature is null_text_inversion_step(embedding, grads, lr). Return the final embedding; an empty grads list leaves it unchanged.

Hint

Null-text inversion optimizes the unconditional embedding so a real image can be reconstructed.

Test cases

case 1

in: [[1,2],[[0.1,0.2]],0.5]

expected: [0.95,1.9]

case 2

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

expected: [-0.75]

case 3

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

expected: [1,0]

case 4

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

expected: [-1]

Starter codePython

def null_text_inversion_step(embedding, grads, lr):
    # Your code here
    pass

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