MTTN: Multi-Pair Text to Text Narratives for Prompt Generation

TL;DR

This paper introduces MTTN, a large-scale dataset with over 12 million text pairs across five stages, designed to improve prompt generation for diffusion models by reflecting real-world internet language and increasing complexity.

Contribution

The creation of the MTTN dataset, which is the largest and most diverse dataset for text-to-text narrative prompts, with a multi-stage design to enhance generative prompt quality.

Findings

MTTN contains over 12 million pairs across 5 stages.

The dataset reflects global internet language usage.

Enhanced prompt complexity improves diffusion model outputs.

Abstract

The increased interest in diffusion models has opened up opportunities for advancements in generative text modeling. These models can produce impressive images when given a well-crafted prompt, but creating a powerful or meaningful prompt can be hit-or-miss. To address this, we have created a large-scale dataset that is derived and synthesized from real prompts and indexed with popular image-text datasets such as MS-COCO and Flickr. We have also implemented stages that gradually reduce context and increase complexity, which will further enhance the output due to the complex annotations created. The dataset, called MTTN, includes over 2.4 million sentences divided into 5 stages, resulting in a total of over 12 million pairs, and a vocabulary of over 300,000 unique words, providing ample variation. The original 2.4 million pairs are designed to reflect the way language is used on the internet globally, making the dataset more robust for any model trained on it.