EasyGen: Easing Multimodal Generation with BiDiffuser and LLMs

TL;DR

Comprehensive quantitative and qualitative experiments show that EasyGen excels in data-efficient training, high-quality image generation, and extendibility, effectively addressing the challenges in multimodal generation.

Abstract

We present EasyGen, an efficient model designed to enhance multimodal understanding and generation by harnessing the capabilities of diffusion models and large language models (LLMs), Unlike existing multimodal models that predominately depend on encoders like CLIP or ImageBind and need ample amounts of training data to bridge modalities,EasyGen leverages BiDiffuser,a bidirectional conditional diffusion model, to foster more efficient modality interactions. Easygen achieves text generation by training a projection layer linking BiDiffuser and an LLM, and facilities image generation by training an adapter to align the LLM's text space with the BiDiffuser's image space, Comprehensive quantitative and qualitative experiments show that EasyGen excels in data-efficient training, high-quality image generation, and extendibility, effectively addressing the challenges in multimodal generation. The source code is available at https://github.com/zxy556677/EasyGen.

Figures (13)

• Figure 1: Our model EasyGen can understand multimodal inputs and generate multimodal responses, as illustrated by model-generated speech bubbles in grey color, which include both text and images.
• Figure 2: Overview of EasyGen.
• Figure 3: The training of BiDiffuser involves finetuning the denoising transformer U-ViT in UniDiffuser with a joint objective of image-to-text and text-to-image tasks.
• Figure 4: Two different ways of aligning BiDiffuser with LLMs.
• Figure 5: Text-to-image generation by EasyGen. LLM generates the response and description of the image. BiDiffuser generates images based on the description.