Tuxemon Dataset

This dataset contains images of mosnters from The Tuxemon Project - an open source effort for a monster catching game. These image-caption pairs can be used for text-to-image tuning & benchmarking.

• All images in this dataset were downloaded from https://wiki.tuxemon.org/Category:Monster
• Some images were upscaled using SDx4 upscaler & HiDiffusion
• Captions generated with
  • BLIP-large (some were manually modified)
  • GPT-4 Turbo

[!TIP] One can use the mix of captions provided in the dataset to study the impact of quality of the captions on the generated images. Some representative works in this area include RECAP and DALL-E 3.

generate_captions_gpt4.py was used to generate captions with GPT4-Turbo. collate_captions.py was used to collate the captions and prepare the final dataset.

Dataset authors

Linoy Tsaban and Sayak Paul