← All research

AAAI Conference on Artificial Intelligence · 2024

LaViP: Language-Grounded Visual Prompting

Nilakshan Kunananthaseelan, Jing Zhang, Mehrtash Harandi

A parameter-efficient way to adapt vision-language models by grounding visual prompts in language, including black-box settings.

  • Visual prompting
  • Vision-language models
  • Model adaptation
Diagram illustrating language-grounded visual prompting with LaViP.
LaViP reprograms a frozen visual encoder through language-grounded prompts at its input.

Research summary

Abstract

LaViP adapts the visual encoder of a vision-language model by learning prompts at the image input. Language provides semantic grounding for those prompts, helping the model reuse its pretrained representation without modifying or extending the encoder’s parameters.

Core idea

Traditional visual prompting learns an input transformation from visual examples alone. LaViP incorporates class-language information when generating the prompt, connecting the downstream task to the semantic structure already present in the vision-language model.

Capabilities

Because adaptation happens at the input, LaViP can work when the model’s internal parameters are frozen or inaccessible. The method is evaluated in few-shot adaptation, base-to-novel generalisation, and transfer-learning settings across datasets including EuroSAT, UCF101, DTD, and CLEVR.

Why it matters

Language grounding improves the accuracy and speed of visual prompt learning while retaining the efficiency and black-box compatibility of model reprogramming.

Use this work

Citation

Kunananthaseelan, N., Zhang, J., & Harandi, M. (2024). LaViP: Language-Grounded Visual Prompting. Proceedings of the AAAI Conference on Artificial Intelligence, 38(3), 2840–2848.

@inproceedings{kunananthaseelan2024lavip,
  title     = {LaViP: Language-Grounded Visual Prompting},
  author    = {Kunananthaseelan, Nilakshan and Zhang, Jing and Harandi, Mehrtash},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  volume    = {38},
  number    = {3},
  pages     = {2840--2848},
  year      = {2024},
  doi       = {10.1609/aaai.v38i3.28064}
}