Post #2942361

https://untwisting-rope.github.io/ pretty fascinating. how to split RoPE into different frequencies, then share only the stylistic / low frequency part between images, to transfer styles in one shot, without any training. Just like you can do semantic math with textual embeddings (e.g., "king" + "woman" = "queen"), this is another example of something that seems like a vague human concept (style), but that still has a mathematical foundation that can be manipulated if you know the right mapping between concepts.