I think it's an analogy. paella is a Spanish rice dish. Rice cooked with various other ingredients like chicken, seafood, peppers, tomatoes, etc... If an ima

I think it's an analogy. paella is a Spanish rice dish. Rice cooked with various other ingredients like chicken, seafood, peppers, tomatoes, etc... If an image (a 2-d array of pixels) is a plate of rice, standard diffusion models denoise starting from each grain of rice (pixel). If you've ever seen a step-by-step output from a diffusion model, you know what I'm talking about. This model makes use of a CNN (convolutional neural net) to decode tokens from the image. A CNN takes m-by-m sections of elements (i.e. a square of pixels) as input, and translates them into 1d vectors as part of the input to the NN. Taking "chunks" of the image like this allows the net to learn things like edges, shapes, groups of color, etc. You could consider the convolutional samples as "chunks" in the paella: the meat, vegetables, and other goodies that make the dish beloved by many.