Stonkinator: An Automatic Generator of Memetic Images

A meme is a transmission unit that carries an idea or behaviour between human beings. Nowadays this term comes associated with the idea of an image or video from the internet, usually with humorous purposes. In an increasingly active and global society, these memes came as a means of entertainment and communication, and throughout the years, many images were created to encode different meanings.

Combining the growing use of memes with the fast-paced communication through social media and other applications, the presented project was developed to help users create their own meme using a simple tool. We developed Stonkinator, a system capable of generating memes in the Stonks format from a text input, using techniques of natural language processing, image analysis and image blending.

The Stonkinator system is divided into four modules: (i) the text handler; (ii) the image obtainer; (iii) the image analyser; and (iv) the image blender. The first module analyses the input sentence given by the user and extracts a theme or subject, which is sent to the second module. The second module obtains a set of images related to the theme identified by the first module. The third module analyses the images and selects one containing a person and another one to be used as the background. It also resizes the images and creates a mask image to be used in the blending process. Lastly, the fourth module receives the image containing the person, the background and the mask and performs the blending. It also writes the sentence used as input, the retrieved word related to the action, and pastes the character’s head (Stonks’ head) in the person’s head place.

We conducted a user study with the goal of understanding the audience’s blending methods preferences and the viability of the system. The results show that our system is able to create satisfactory images, being the most visually appealing the ones with a simple pasting blending method, and that users would share the generated memes on their social networks or in the context of a private text conversation.

Using the feedback gathered during the user study we improved the system further. One of the improvements focused on the typography, using the Impact font to write captions and a variable font size for the word written in the image. Furthermore, we added a feature that introduced a small typo in the written word to get the outputs closer to the desired results.

The paper was accepted in the 14th International Conference on Computational Creativity, ICCC’23 and will be made available soon.