The process here called “unbundling” is, I think, a type of metaprompt. It turns a sending into a summoning.
When it comes to the world of Multimodal AI models like DALL-E 2, I would define Unbundling as:
The process of exploding a label of an artistic style, known work, art movement, or universally recognized symbol into a longer description of its various characteristics and components.
(from https://bakztfuture.substack.com/p/dall-e-2-unbundling)
The idea is to dismantle a specific invocation, like “Van Gogh starry night”, into its component parts: style, mood, material, &al. Then use this phrase to call upon that energy without specifying its true name. The author uses GPT-3 to describe concepts “in the zeitgeist”.
This is like if I were to say, “I am calling into myself aspects of the messenger, the merchant, the trickster, the traveler, the thief” but never to say the name Hermes.
(If I were to do magic, which I definitely don’t. Anymore. Even in exceptional circumstances.)
One could imagine also getting the embedding of this concept, and multiplying or adding it to your prompt vector to incorporate the style. We already do this in many models with the pipe indicator ”|”, which combines and weights prompts.
Can these vectors compose? Could we use them to create scenes? Can we compose them like functions?
This is kind of already possible… enables a wicked game of paper telephone?
Feel like the dream tool for this would be a reversal of the model's basic operation: I'd like to provide a set of Savador Dali paintings, and recieve a string of near-enough-to-natural-language text reflecting what they had in common, that would produce the same style fed back.
— 👁👃👁 (@waxreplica) July 7, 2022
Using something like personality-clip for instance
Personality-CLIP: generating image captions with personalityhttps://t.co/jxoKWsnUfs
— dzryk (@dzryk) August 25, 2021