Latent-Tensor Editing: A First Look at Noise Gradients

Overview:

I have output for 8 image seeds with noise-tensor-editing. For each we have an image-pair and the progress of p_hat_release. I am currently allowing all 15 noise-tensors to be edited. This is also with full male and race orthogonalization.

Output:

• Allowing for noise tensor editing currently creates artifacts in the images (very evident in seed-100 and seed-107)
• Gradient descent is not progressing monotonically (you can see the shape of the p_hat_release curves)
• We can see that the range of p_hat_release appears to be larger though, since we are actually seeing differences of about 35-45 %-points

Plan of Attack:

• I will produce the same image seeds with a reduced step-size to see whether that solves the gradient-descent problem
• I will also restrict the noise tensors that it can edit systematically
• My thinking is that we should keep lower-dim noise tensors static (the first 4-5 layers) and allow the final few to change
  • Lower-dim noise tensors impact larger features
  • Noise infused later in the model should impact more minute features
• If none of this works, we may need to think of including a gradient-penalty to restrict the artifacts being created

Seed 100

Seed 101

Seed 102

Seed 103

Seed 104

Seed 105

Seed 106

Seed 107