ELITE is out! We need this as an extension.

[bbecausereasonss]

https://github.com/csyxwei/ELITE

Given an image indicates the target concept (usually an object), we propose a learning-based encoder ELITE to encode the visual concept into the textual embeddings, which can be further flexibly composed into new scenes. It consists of two modules: (a) a global mapping network is first trained to encode a concept image into multiple textual word embeddings, where one primary word (w0) for well-editable concept and other auxiliary words (w1···N) to exclude irrelevant disturbances. (b) A local mapping network is further trained, which projects the foreground object into textual feature space to provide local details.

https://github.com/csyxwei/ELITE/raw/main/assets/results.png