Unifying Multimodal Transformer for Bi-directional Image and Text Generation

TL;DR

This paper introduces a unified Transformer-based model for bi-directional image and text generation, effectively handling both tasks with a single architecture and improving performance metrics on MS-COCO.

Contribution

The work presents a novel unified multimodal Transformer framework that jointly learns image-to-text and text-to-image generation, simplifying design and enhancing results.

Findings

Significant FID reduction from 37.0 to 29.9 in text-to-image generation.

CIDEr-D score improvement from 100.9% to 122.6% in image-to-text generation.

Effective sequence-level training with two-level granularity features.

Abstract

We study the joint learning of image-to-text and text-to-image generations, which are naturally bi-directional tasks. Typical existing works design two separate task-specific models for each task, which impose expensive design efforts. In this work, we propose a unified image-and-text generative framework based on a single multimodal model to jointly study the bi-directional tasks. We adopt Transformer as our unified architecture for its strong performance and task-agnostic design. Specifically, we formulate both tasks as sequence generation tasks, where we represent images and text as unified sequences of tokens, and the Transformer learns multimodal interactions to generate sequences. We further propose two-level granularity feature representations and sequence-level training to improve the Transformer-based unified framework. Experiments show that our approach significantly improves previous Transformer-based model X-LXMERT's FID from 37.0 to 29.9 (lower is better) for text-to-image generation, and improves CIDEr-D score from 100.9% to 122.6% for fine-tuned image-to-text generation on the MS-COCO dataset. Our code is available online.