Now manipulating features directly surfaced from the image embedding model. In a way its like sifting through the mind of the model, nudging certain parts to guide toward a particular direction and then re-generating the image.

Experimenting with auto "vibe" discovery for images using gpt-4o and then increasing/decreasing strength by generating images using interpolated embeddings.

My current mental model - asking gpt4o about the image vs. interpreting features from the model is like asking what a third party thinks about a painting vs. reading the mind of the painter. A difference in level/flexibility of control when steering re-generations.

Tinkering. Co-Founder @ Livepeer. https://www.yondonfu.com/

My current mental model - asking gpt4o about the image vs. interpreting features from the model is like asking what a third party thinks about a painting vs. reading the mind of the painter. A difference in level/flexibility of control when steering re-generations.

https://warpcast.com/yondon.eth/0xd2ad34d7