🎨 ShowFlow: From Robust Single Concept
to Condition-Free Multi-Concept Generation

1 University of Science, VNU-HCM, Ho Chi Minh City, Vietnam
2 Vietnam National University, Ho Chi Minh City, Vietnam
3 Monash University, Melbourne, Australia
4 University of Dayton, Ohio, United States
* Equal contribution; † Corresponding author

Teaser

Abstract

Customizing image generation remains a core challenge in controllable image synthesis. For single-concept generation, maintaining both identity preservation and prompt alignment is challenging. In multi-concept scenarios, relying solely on a prompt without additional conditions like layout boxes or semantic masks, often leads to identity loss and concept omission. In this paper, we introduce ShowFlow, a comprehensive framework designed to tackle these challenges. We propose ShowFlow-S for single-concept image generation, and ShowFlow-M for handling multiple concepts. ShowFlow-S introduces a KronA-WED adapter, which integrates a Kronecker adapter with weight and embedding decomposition, and employs a disentangled learning approach with a novel attention regularization objective to enhance single-concept generation. Building on this foundation, ShowFlow-M directly reuses the learned models from ShowFlow-S to support multi-concept generation without extra conditions, incorporating a Subject-Adaptive Matching Attention (SAMA) and a layout consistency strategy as the plug-and-play module. Extensive experiments and user studies validate ShowFlow’s effectiveness, highlighting its potential in real-world applications like advertising and virtual dressing.

Method Overview

ShowFlow-S for single-concept learning and generation

We introduce a new adapter called Kronecker Adaptation with Weight and Embedding Decomposition (KronA-WED) and a fine-tuning strategy that combines disentangled learning with a novel attention regularization objective to achieving the balance of reconstruction and editability.

ShowFlow-S Overview

ShowFlow-M for condition-free multi-concept generation

Given $N$ learned concepts using KronA-WEDs, we adopt gradient fusion to obtain the fused update weights. Using this weight for generating multi-concept images, we propose Subject-Adaptive Matching Attention (SAMA) module to preserve the identity details of each concept within the composite image, and layout consistency guidance to mitigate the concepts missing issue.

ShowFlow-M Overview

ShowFlow-M Layout Consistency Overview

📈 Quantitative Results

ShowFlow-S component exhibit comparable results in all metrics, thereby demonstrating its capability in balancing the trade-off of reconstruction and editability. Meanwhile, ShowFlow-M outperforms other methods in identity preservation and layout consistency.

Single-concept Generation

Single-concept generation

Condition-Free Multi-concept Generation

Condition-Free multi-concept generation

👀 Qualitative Results

Single-concept Generation

Single-concept generation

Condition-Free Multi-concept Generation

Condition-Free multi-concept generation

Citation

Feel free to contact Trong-Vu Hoang or Quang-Binh Nguyen for any question. If you find this work useful, please consider citing:


          @misc{hoang2025showflowrobustsingleconcept,
            title={ShowFlow: From Robust Single Concept to Condition-Free Multi-Concept Generation}, 
            author={Trong-Vu Hoang and Quang-Binh Nguyen and Thanh-Toan Do and Tam V. Nguyen and Minh-Triet Tran and Trung-Nghia Le},
            year={2025},
            eprint={2506.18493},
            archivePrefix={arXiv},
            primaryClass={cs.CV},
            url={https://arxiv.org/abs/2506.18493}, 
          }