PictSure 0.2.2

A package for generalized image classification using In-Context-Learning with PyTorch.

PictSure: In-Context Learning for Image Classification

PictSure is a deep learning library designed for in-context learning using images and labels. It allows users to provide a set of labeled reference images and then predict labels for new images based on those references. This approach eliminates the need for traditional training, making it highly adaptable for various classification tasks.

Features

• In-Context Learning: Predict labels for new images using a set of reference images without traditional model training.
• Multiple Model Architectures: Choose between ResNet and ViT-based models for your specific needs.
• Pretrained Models: Use our pretrained models or train your own.
• Torch Compatibility: Fully integrated with PyTorch, supporting CPU and GPU.
• Easy-to-use CLI: Manage models and weights through a simple command-line interface.

Installation

pip install PictSure

Quick Start

from PictSure import PictSure
import torch

DEVICE = "cpu" # or cuda, mps

model = PictSure.from_pretrained("pictsure/pictsure-vit")
model = model.to(DEVICE)

# Set your reference images and labels
model.set_context_images(reference_images, reference_labels)

# Make predictions on new images
predictions = model.predict(new_images)

Examples

For a complete working example, check out the Jupyter notebook in the Examples directory:

Examples/example.ipynb

This notebook demonstrates:

• Model initialization
• Loading and preprocessing images
• Setting up reference images
• Making predictions
• Visualizing results

Citation

If you use this work, please cite it using the following BibTeX entry:

@article{schiesser2025pictsure,
  title={PictSure: Pretraining Embeddings Matters for In-Context Learning Image Classifiers},
  author={Schiesser, Lukas and Wolff, Cornelius and Haas, Sophie and Pukrop, Simon},
  journal={arXiv preprint arXiv:2506.14842},
  year={2025}
}

License

This project is open-source under the MIT License.

Contributing

Contributions and suggestions are welcome! Open an issue or submit a pull request.

Contact

For questions or support, open an issue on GitHub.