Comment by spywaregorilla

Is there any way to approximate embeddings for a novel image?

novel meaning user provided, not generated by the model or in the training set.