Multimodal Robust Prompt Distillation for 3D Point Cloud Models

Official Code for Papaer in AAAI2026 Multimodal Robust Prompt Distillation for 3D Point Cloud Models

paper here https://arxiv.org/abs/2511.21574v1

Cheers! We have been accepted!

Will update CKPTs soon, You can directly contact XiangGu2003@icloud.com XiangGu2003@connect.hku.hk if you have code-problems or just open issues here. I am glad to help. Meanwhile i am going to polish and update a more re-produceable version.

This repository contains the official PyTorch implementation for the experiments in our paper, "Multimodal Robust Prompt Distillation for 3D Point Cloud Models".

Please Note: This repository currently provides the raw code used for our experiments to ensure full reproducibility for the review process. We are committed to releasing a refactored, fully-documented, and user-friendly version of the code upon the acceptance of our paper.

Environment Setup

Our code is built upon the experimental environment of the Uni3D baseline.

Reproduction Steps

if you want to prompt distill under uni3d,pls follow https://github.com/baaivision/Uni3D otherwise pls choose the model you desire

1. Data Preparation

• Datasets: Please download the ModelNet40 and ScanObjectNN datasets. You may need to specify the path to your datasets in the training/testing scripts.
• Attack Data Generation: For a fair and direct comparison with prior work, our adversarial attack generation protocol strictly follows the public benchmark established by methods like AOF (https://github.com/code-roamer/AOF). We utilize their methodologies for generating all adversarial samples.
  • Our script test.py includes functionalities for generating attack data.
  • We have only made minor adjustments to the data format to align with our processing pipeline; the core attack logic remains identical to ensure fairness.

2. Training

To train our MRPD model from scratch, run the train_multi_prompt.py script. This script handles the multi-teacher distillation and prompt learning process.

python train_multi_prompt.py #params can be seen in the file.