---
license: apache-2.0
library_name: transformers
base_model: DAMO-NLP-SG/VideoLLaMA3-7B
tags:
  - multimodal
  - video-llm
  - video-understanding
  - video-grounding
  - spatio-temporal-grounding
  - temporal-grounding
  - referring-expression-comprehension
  - grounding
  - reasoning
  - custom_code
  - pytorch
---

<div align="center">

# DeViL-7B

**Official checkpoint for "Detector-Empowered Video Large Language Model for Efficient Spatio-Temporal Grounding"**

[Paper](https://arxiv.org/abs/2512.06673) | [Code](https://github.com/gaostar123/DeViL)

</div>

<p align="center">
  <img src="https://raw.githubusercontent.com/gaostar123/DeViL/main/assets/intro_v10.png" alt="DeViL teaser" width="100%">
</p>

## Overview

DeViL is a detector-empowered video large language model designed for efficient spatio-temporal video grounding (STVG) and grounded video reasoning. Instead of relying on long autoregressive coordinate decoding or expensive candidate construction, DeViL offloads dense spatial grounding to a fully parallel detector. It distills the user query into a detector-compatible reference-semantic token and uses temporal consistency regularization to maintain object coherence across frames.

This repository hosts the official DeViL-7B checkpoint released by the authors.

## Highlights

- Detector-empowered grounding for efficient spatio-temporal localization
- Strong performance reported in the paper: 43.1 m_vIoU on HC-STVG and 14.33 FPS
- Preserves the backbone MLLM's general video understanding and reasoning ability
- Supports both image and video inputs in the official demo pipeline

## Links

- Paper: https://arxiv.org/abs/2512.06673
- Code: https://github.com/gaostar123/DeViL