Novel Object Synthesis via Adaptive Text-Image Harmony

TL;DR

This paper introduces Adaptive Text-Image Harmony (ATIH), a novel method that balances text and image features in diffusion models to synthesize new objects that harmoniously combine input text and images.

Contribution

We propose a new adaptive method with scale factors, balanced loss, and similarity scoring to improve object synthesis from text and images in diffusion models.

Findings

Enhanced object synthesis quality demonstrated in experiments

Effective balancing of text and image features achieved

Generated objects exhibit high fidelity and editability

Abstract

In this paper, we study an object synthesis task that combines an object text with an object image to create a new object image. However, most diffusion models struggle with this task, \textit{i.e.}, often generating an object that predominantly reflects either the text or the image due to an imbalance between their inputs. To address this issue, we propose a simple yet effective method called Adaptive Text-Image Harmony (ATIH) to generate novel and surprising objects. First, we introduce a scale factor and an injection step to balance text and image features in cross-attention and to preserve image information in self-attention during the text-image inversion diffusion process, respectively. Second, to better integrate object text and image, we design a balanced loss function with a noise parameter, ensuring both optimal editability and fidelity of the object image. Third, to adaptively adjust these parameters, we present a novel similarity score function that not only maximizes the similarities between the generated object image and the input text/image but also balances these similarities to harmonize text and image integration. Extensive experiments demonstrate the effectiveness of our approach, showcasing remarkable object creations such as colobus-glass jar. Project page: https://xzr52.github.io/ATIH/.