ConceptFlow: Unified Framework for Personalized Image Generation
Keywords: Personalized Image Generation, Visual Guidance
Abstract: Personalized image generation is an appealing area of research within controllable image generation due to its diverse potential applications. Despite notable advancements, generating images based on single or multiple concepts remains challenging. For single-concept generation, it is difficult to strike a balance between identity preservation and prompt alignment, especially in complex prompts. When it comes to multiple concepts, creating images from a single prompt without extra conditions, such as layout boxes or semantic masks, is problematic due to significantly identity loss and concept omission. In this paper, we introduce ConceptFlow, a comprehensive framework designed to tackle these challenges. Specifically, we propose ConceptFlow-S and ConceptFlow-M for single-concept generation and multiple-concept generation, respectively. ConceptFlow-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. On the other hand, ConceptFlow-M leverages models learned from ConceptFlow-S to directly generate multi-concept images without needed of additional conditions, proposing Subject-Adaptive Matching Attention (SAMA) module and layout consistency guidance strategy. Our extensive experiments and user study show that ConceptFlow effectively addresses the aforementioned issues, enabling its application in various real-world scenarios such as advertising and garment try-on.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4581
Loading