FiGMaQ
This dataset consists of a quintuple format in the form of <reference image, reference caption, modification text, target image, target caption>. Unlike traditional LLM-generated multimodal triplet datasets, our dataset introduces three key characteristics that distinguish it:
- Detailed Image Captions: Each image is accompanied by a detailed caption that elaborates on various attributes of the image. These captions are typically longer than 100 tokens, offering a nuanced and rich description of the visual content, which enhances multimodal understanding.
- Rich Modification Text: The modification text in this dataset provides a more precise description of changes made to the image. It is written in a more natural, human-like style, incorporating vague and imprecise terms, making it closer to how humans would describe adjustments or edits in everyday language.
- Quintuple Structure: Unlike typical triplets, each sample in this dataset consists of five parts. This expanded format facilitates a wide range of fine-tuning tasks, including multimodal generation and retrieval, enabling diverse applications that require an integrated understanding of both images and text.
We will release our data and code soon!