/ic/ - Artwork/Critique

>>6473854
Yea basically, the patterns are clusters inside the latent space. The way these diffusion models work is they take some random noise, then try to "remove" that noise step by step in order to "reconstruct" an image and the resulting images will be interpolations from the latent space (which was constructed from the LAION dataset during the training process). Typing prompts basically points the algorithm to certain areas in the latent space (so if you type "cat" it will try to interpolate something from the "cat" cluster). Check out this article which explains latent space interpolation in more detail and shows how it differs from "regular" interpolation in the pixel space:
>https://hackernoon.com/latent-space-visualization-deep-learning-bits-2-bd09a46920df

>>6473860
Ok check out this OC I made