Infusion: Preventing Customized Text-to-Image Diffusion from Overfitting

Zeng Weili, Yichao Yan, Qi Zhu, Zhuo Chen, Pengzhi Chu, Weiming Zhao, Xiaokang Yang

Abstract:

Text-to-image (T2I) customization aims to create images that embody specific visual concepts delineated in textual descriptions.
However, existing works still face a main challenge, \textbf{concept overfitting}.
To tackle this challenge, we first analyze overfitting, categorizing it into concept-agnostic overfitting, which undermines non-customized concept knowledge, and concept-specific overfitting, which is confined to customize on limited modalities, \ie, backgrounds, layouts, styles. To evaluate the overfitting degree, we further introduce two metrics, \ie, Latent Fisher divergence and Wasserstein metric to measure the distribution changes of non-customized and customized concept respectively.
Drawing from the analysis, we propose Infusion, a T2I customization method that enables the learning of target concepts to avoid being constrained by limited training modalities, while preserving non-customized knowledge. Remarkably, Infusion achieves this feat with remarkable efficiency, requiring a mere \textbf{11KB} of trained parameters.
Extensive experiments also demonstrate that our approach outperforms state-of-the-art methods in both single and multi-concept customized generation.

Primary Subject Area: [Generation] Generative Multimedia

Secondary Subject Area: [Content] Vision and Language

Relevance To Conference: Our work is dedicated to contributing to the field of multimodal processing, with a focus on solving the overfitting problem of current customized text to image generation, allowing users to better generate customized content based on language guidance.

Submission Number: 803