MIGC: Multi-Instance Generation Controller for Text-to-Image Synthesis

TL;DR

This paper introduces the Multi-Instance Generation Controller (MIGC), a novel method for generating multiple diverse instances within a single image with precise control over location and attributes, advancing text-to-image synthesis.

Contribution

The paper proposes MIGC, an innovative approach that decomposes multi-instance generation into subtasks with an attention mechanism, and introduces the COCO-MIG benchmark for evaluation.

Findings

MIGC achieves high control accuracy in quantity, position, and attributes.

Extensive experiments demonstrate MIGC's superior performance on COCO-MIG and other benchmarks.

The approach enhances the versatility of text-to-image synthesis models.

Abstract

We present a Multi-Instance Generation (MIG) task, simultaneously generating multiple instances with diverse controls in one image. Given a set of predefined coordinates and their corresponding descriptions, the task is to ensure that generated instances are accurately at the designated locations and that all instances' attributes adhere to their corresponding description. This broadens the scope of current research on Single-instance generation, elevating it to a more versatile and practical dimension. Inspired by the idea of divide and conquer, we introduce an innovative approach named Multi-Instance Generation Controller (MIGC) to address the challenges of the MIG task. Initially, we break down the MIG task into several subtasks, each involving the shading of a single instance. To ensure precise shading for each instance, we introduce an instance enhancement attention mechanism. Lastly, we aggregate all the shaded instances to provide the necessary information for accurately generating multiple instances in stable diffusion (SD). To evaluate how well generation models perform on the MIG task, we provide a COCO-MIG benchmark along with an evaluation pipeline. Extensive experiments were conducted on the proposed COCO-MIG benchmark, as well as on various commonly used benchmarks. The evaluation results illustrate the exceptional control capabilities of our model in terms of quantity, position, attribute, and interaction. Code and demos will be released at https://migcproject.github.io/.