Test-Time Visual In-Context Tuning

Jiahao Xie^1,2  Alessio Tonioni³  Nathalie Rauschmayr³  Federico Tombari³  Bernt Schiele^1,2

¹Max Planck Institute for Informatics  ²VIA Research Center  ³Google

CVPR 2025

We present VICT, a test-time visual in-context tuning method that can adapt visual in-context learning models on the fly with a single test sample. VICT can be applied to a wide range of unseen domains and tasks at test time.

:open_book: For more results, please refer to our paper

---

📣 News

• [03/2025] 🔥 VICT is released on arXiv.

🌟 Method

VICT is a simple yet effective test-time training approach to adapt visual in-context learning (VICL) models on the fly. The motivation is that each test input offers a hint about the test distribution. Thus, we modify a VICL model at test time to make full use of this hint by setting up a one-sample learning problem.

Specifically, we flip the role between the task prompts and the test sample and use a cycle consistency self-supervised loss to reconstruct the original task prompt output. Our key insight is that a model should be aware of a new test distribution if it can successfully recover the original task prompts.

🤗 Qualitative Examples

Unseen Domains

Middle-/High-Level Tasks with Corruptions

Low-Level Tasks with Corruptions

Unseen Tasks

🛠️ Usage

Installation

See installation instructions.

Data

See data instructions.

Training

See training instructions.

Evaluation

See evaluation instructions.

👨‍💻 Todo

• Release the arXiv version.
• Release the code.

📘 Citation

If you find this work useful for your research, please consider citing our paper:

@inproceedings{xie2025test,
  title = {Test-Time Visual In-Context Tuning},
  author = {Xie, Jiahao and Tonioni, Alessio and Rauschmayr, Nathalie and Tombari, Federico and Schiele, Bernt},
  booktitle={CVPR},
  year = {2025}
}

❤️ Acknowledgement

We acknowledge the use of the following public code in this project: Painter, MAE, BEiT, detectron2, Mask2Former, bts, mmcv, mmdetetection, mmpose, MIRNet, MPRNet, and Uformer.