search

sail-sg / CLoT

CVPR'24, Official Codebase of our Paper: "Let's Think Outside the Box: Exploring Leap-of-Thought in Large Language Models with Creative Humor Generation".
https://zhongshsh.github.io/CLoT
264 stars 11 forks source link

not funny to general images #12

Open DietDietDiet opened 3 months ago

DietDietDiet commented 3 months ago
image

I can rarely find funny responses, can it generalized well to a broader range of images?

zhongshsh commented 3 months ago

Creativity is challenging. The images used in the Oogiri game are general images. In fact, as mentioned in our paper, innovation itself is a tough endeavor. Even top human players find it challenging to come up with responses for general images in the Oogiri game (such as video: link 19:42~28:57). Therefore, it's hard to expect models to generate satisfying responses for any given image. Despite there may exist images where finding a funny response is nearly impossible, the creative humorous responses we've seen so far demonstrate the effectiveness of our model.

Humor is so subjective. For example, your generated response "最近一直加班，每天晚上都在加班" is a dark humor that thinks outside the box for me while maybe you don't think it funny. I think it fits well with the image of deserted streets and the description of someone looking weary from frequent late-night work. We suggest trying out more generations to find responses that suit your preferences.

I've also made some attempts for your image, and here are the results. At least these model outputs are what I think interesting.

example