Learning Identifiable Concepts for Compositional Image Generation

Shaoan Xie, Yujia Zheng, Ignavier Ng, Kun Zhang

Keywords: concept; composition; image generation

TL;DR: We present a method for compositional image generation and editing with identifiability guarantees.

Abstract:

Humans have the ability to decompose objects into parts and relationships and
create new objects by properly combining existing concepts. However, enabling
machines to achieve this in real-world tasks remains a challenge. In this paper,
we investigate how to teach machines compositional image generation through
learning identifiable concepts. To derive concepts from attribute labels, we formulate the minimal change principle and propose a method to limit the information introduced by each label. Additionally, to address dependent attribute labels
(with causal influences in between or common causes behind them), we present
a causal conditioning approach to disentangle concepts from these correlations.
Our framework enhances data efficiency, interpretability, and control, while enabling sampling from unseen combinations. We validate our method on various
compositional image generation and editing tasks, demonstrating its effectiveness
through superior performance.

Primary Area: generative models

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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 8486