This is the official repository for the CVPR 2026 paper, Learning to Select Visual In-Context Demonstrations.

Our framework, Learning to Select Demonstrations (LSD), reframes K-shot demonstration selection as a sequential decision-making problem. We utilize a Dueling Deep Q-Network (DQN) agent with a query-centric Transformer Decoder to actively balance visual relevance with diversity, constructing optimal demonstration sets for visual regression tasks.

---

Setup

cd Learning-to-Select-Demonstrations
conda create -n LSD python=3.10 -y
conda activate LSD
pip install -r requirements.txt

---

Process Data

1. Download Dataset

Download the datasets from the following sources:

• Age Prediction (UTKFace) https://susanqq.github.io/UTKFace/
• Aesthetic Score (AVA) https://www.kaggle.com/datasets/nicolacarrassi/ava-aesthetic-visual-assessment
• Facial Beauty (SCUT-FBP5500 v2) https://www.kaggle.com/datasets/pranavchandane/scut-fbp5500-v2-facial-beauty-scores
• Wild Image Quality (KonIQ-10k) https://database.mmsp-kn.de/koniq-10k-database.html
• Modified Image Quality (KADID-10k) https://database.mmsp-kn.de/kadid-10k-database.html

2. Prepare Dataset

./scripts/prepare_data.sh

3. Update datasets.yaml

---

LSD Training

./scripts/train_lsd.sh

Note: Results will be saved in the train_res directory.

LSD Evaluation

./scripts/eval_lsd.sh

Note: Results will be saved in the eval_res directory.

---

Citation

If you find this work helpful or use our code in your research, please consider citing our paper:

@inproceedings{lee2026learning,
  title={Learning to Select Visual In-Context Demonstrations},
  author={Lee, Eugene and Lin, Yu-Chi and Diao, Jiajie},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={9455--9465},
  year={2026}
}