Two anchor fields A and B differ by one edit, such as a blur, a stretch or a split. The direction between their embeddings, η = ζmB − ζmA, is that edit read off the latent space. It is then added, with a coefficient c from −1 to 2, to the embedding of a completely different field, yours, and the result is decoded: c = 1 applies the edit once, a negative c reverses it.
Checking the server…Each preset comes from two analytic fields that differ by one operation, such as a sharp and a broad bump for diffusion. Their single-partition embeddings ζmA and ζmB define the direction η = ζmB − ζmA.
The input is encoded as a single partition, like the anchors: its shape becomes one embedding ζm, its position, extent and value range the static frame.
The embedding ζm becomes |ζm| (ζm + c η) / |ζm + c η|: it moves along η and stays on the sphere. The static frame is kept.
The input is decoded for coefficients c from −1 to 2: c = 0 is your input as one embedding, a negative c reverses the edit.
The input is one partition, so the direction acts on its overall shape, as on the anchors it was measured on; a detailed input is reconstructed coarsely by a single embedding.