Diffusion Momentum

[tom-doerr]

I think the human-in-the-loop part could be much enhanced by adding momentum to the diffusion phase. That way it might be much easier to get significant and meaningful changes in the images.
As pseudocode:

def diffusion(image):
    momentum = 0
    while True:
        diffused_images = diffuse(image.latent_encoding + momentum)
        new_image = user_choice(diffused_images)
        momentum += new_image.latent_encoding - image.latent_encoding
        if stop_diffusion:
            break

Any chance we could get something like that or is it already possible to access the latent encoding so I can implement it myself?