Compositional Text-to-Image Generation with Dense Blob Representations

TL;DR

This paper introduces BlobGEN, a novel compositional text-to-image generation model that uses dense blob representations for fine-grained, controllable scene synthesis, leveraging large language models for improved prompt understanding.

Contribution

The paper proposes dense blob representations for scene decomposition, a blob-grounded diffusion model, and an in-context learning approach with LLMs for enhanced compositional generation.

Findings

Achieves superior zero-shot generation quality on MS-COCO

Demonstrates better layout-guided controllability

Exhibits improved numerical and spatial correctness on benchmarks

Abstract

Existing text-to-image models struggle to follow complex text prompts, raising the need for extra grounding inputs for better controllability. In this work, we propose to decompose a scene into visual primitives - denoted as dense blob representations - that contain fine-grained details of the scene while being modular, human-interpretable, and easy-to-construct. Based on blob representations, we develop a blob-grounded text-to-image diffusion model, termed BlobGEN, for compositional generation. Particularly, we introduce a new masked cross-attention module to disentangle the fusion between blob representations and visual features. To leverage the compositionality of large language models (LLMs), we introduce a new in-context learning approach to generate blob representations from text prompts. Our extensive experiments show that BlobGEN achieves superior zero-shot generation quality and better layout-guided controllability on MS-COCO. When augmented by LLMs, our method exhibits superior numerical and spatial correctness on compositional image generation benchmarks. Project page: https://blobgen-2d.github.io.