ML Tarot Project


The booklet for the deck can be found here. You can learn more about this project in my DIS 2022 pictorial, co-authored with Daniela Rosner:

Caitlin Lustig and Daniela Rosner. “From Explainability to Ineffability? ML Tarot and the Possibility of Inspiriting Design”. Conference on Designing Interactive Systems. ACM, 2022. [pdf]

Creation (text lightly edited from the DIS pictorial)

To create the ML Tarot deck, I engaged in a multi-stage process which involved using ML (machine learning) to generate descriptions of the cards and then, using another ML algorithm, using those descriptions to generate the images on the cards.

I began by generating descriptions of the cards to put in a booklet. Tarot decks are often accompanied by booklets that help readers interpret the cards. To generate the descriptions, I used GPT-2 (using a copy of a Colab notebook), a large language model trained on a large corpus of web pages, which can generate human-like text. To use GPT-2, the ML algorithm must first be provided some text, and it will then use this text to predict what words should follow. I provided it with short lightly edited descriptions of cards from These sentences from included information about the traditional imagery on the cards and/or the traditional meaning of the cards. GPT-2 used these sentences to generate longer texts. For each card, I used GPT-2 to generate multiple descriptions (in total, about 132 thousand words) and then for each card, I selected the description with the most “imaginative” imagery, narrative, or writing style.

To generate the images shown on each card, I fed key words or quotes from these descriptions into using their VQGAN+CLIP algorithm. This algorithm can generate images based from text prompts. In some cases, I generated multiple images before picking ones that were the most evocative and representative of the textual aesthetic of the descriptions. Lastly, I created a booklet, linked above, with the ML-generated descriptions and printed cards with the ML-generated images.