| pipeline_tag: image-text-to-text | |
| # SPLASH-1B: Wake up for Touch! Mask-isolated Tactile Alignment Learning in MLLMs | |
| This repository contains the SPLASH-1B model checkpoint presented in the paper [Wake up for Touch! Mask-isolated Tactile Alignment Learning in MLLMs](https://huggingface.co/papers/2607.00302). | |
| SPLASH-1B integrates tactile perception into vision-language models by using [InternVL2.5-1B](https://huggingface.co/OpenGVLab/InternVL2_5-1B) as the base MLLM backbone and a ViT-Tiny tactile frontend. | |
| - **Project Page:** [https://ewha-mmai.github.io/splash/](https://ewha-mmai.github.io/splash/) | |
| - **Repository:** [https://github.com/ewha-mmai/splash](https://github.com/ewha-mmai/splash) | |
| ## Description | |
| SPLASH integrates tactile perception into vision-language models through a two-stage pipeline: | |
| 1. **Dormant Mask Generation:** Quantifies the significance of each pretrained parameter via weight & activation importance scoring on the LLM backbone, partitioning the parameter space into a dormant and critical subspace. | |
| 2. **Mask-Guided Sparse Training:** Updates the isolated dormant subspace to align tactile representations while freezing the critical subspace to safeguard established vision-language reasoning, preventing catastrophic forgetting. | |
| <p align="center"> | |
| <img src="https://raw.githubusercontent.com/ewha-mmai/splash/main/assets/figure.png" width="100%"> | |
| </p> | |
| ## Citation | |
| If you find SPLASH useful, please cite the paper: | |
| ```bibtex | |
| @inproceedings{park2026splash, | |
| title={Wake up for Touch! Mask-isolated Tactile Alignment Learning in MLLMs}, | |
| author={Yoonhyung Park and Minji Kim and Sungwon Moon and Jiyoung Lee}, | |
| booktitle={European Conference on Computer Vision (ECCV)}, | |
| year={2026} | |
| } | |
| ``` |