Foundation Vision Models are Unsupervised Image Canonicalizers

Utkarsh Singhal, Ryan Feng, Stella X. Yu, Atul Prakash

Keywords: Invariance, Canonicalization, Foundation Models, CLIP, SD, Augmentation, Vision, Robustness

TL;DR: A training-free and dataset/model agnostic method that uses foundation models to undo transformations such as rotation, lighting, and viewpoint shifts -- thus achieving invariance.

Abstract: One of the most significant and longstanding problems in computer vision is invariance - the ability to robustly handle changes in real-world transformations such as rotation, viewpoint, and lighting. Unfortunately, popular foundation models remain brittle under such transformations. While existing solutions towards invariance have shown promise, they all fundamentally require some model training, limiting their ability to adapt broadly to new tasks, transformations, and datasets. Our key insight is that foundation model priors can be used to reason about transformations. We thus propose Foundation Model Canonicalization (FMC), an approach that can undo nuisance transformations in images without any model training. With a single core approach, FMC can make models like CLIP and SAM invariant to different transformations without any training or fine-tuning. Our approach FMC flexibly adapts to new foundation models and tasks, making it significantly easier for newer and larger models to achieve invariance.

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.

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: 11967