[PYTHON] Papers: Papers that create various images just by specifying words such as "red" + "round" (SCAN: Learning Hierarchical Compositional Visual Concepts)

About the mechanism

Understand the meaning from the image. A large number of apple images → to the image. As a result, features such as round, red, and small are acquired. At the time of materialization, an image is generated based on the acquired concept. There are restrictions such as round, red, and small, but there are no other restrictions. Therefore, since there is a degree of freedom in lighting and background, a variety of images are generated.

Super important point

It can be reached by the following instructions. "blue" AND "small" (down the hierarchy from the more general to the more specific), "blueeberry" IN COMMON "blueebell" (up the hierarchy from the more specific to the more general) You can reach a new node corresponding to the concept {blue, small} by instructing (to let) or "blueeberry" IGNORE "round" (also to raise the hierarchy).

https://arxiv.org/pdf/1707.03389.pdf

Recommended Posts

Papers: Papers that create various images just by specifying words such as "red" + "round" (SCAN: Learning Hierarchical Compositional Visual Concepts)