---
base_model: Qwen/Qwen3-VL-4B-Instruct
license: apache-2.0
library_name: transformers
pipeline_tag: video-text-to-text
tags:
- camera-movement
- video-understanding
- qwen3-vl
- vggt-injection
---

# CamInject-4B

Camera-movement understanding model that **injects frozen VGGT camera tokens** into
`Qwen/Qwen3-VL-4B-Instruct`. Given a video, it outputs structured JSON describing every
camera-movement segment.

- **Paper**: [Temporally Grounded Compositional Camera Motion Understanding via Geometric Knowledge Distillation](https://huggingface.co/papers/2608.10932)
- **Project page**: https://ddz16.github.io/cammotion.github.io
- **Code**: https://github.com/ddz16/CamDistill

> ⚠️ **This model cannot be loaded with plain 🤗 Transformers.** It requires a custom model type
> (registered via a plugin) **and** runs VGGT online to produce camera tokens. Loading it as a
> standard `Qwen3VLForConditionalGeneration` would not work correctly. Use the CamDistill repo.

## Usage

Clone the [CamDistill repo](https://github.com/ddz16/CamDistill) and clone VGGT-Omega (set `VGGT_OMEGA_REPO`, see the repo's
setup). CamInject runs VGGT online during inference:

```bash
VGGT_TEACHER_TYPE=vggt_omega \
python camera_movement_sft/infer_single.py \
    --model ddz16/CamInject-4B \
    --video /path/to/video.mp4 \
    --variant caminject
```

See the repo's README for environment setup and batch evaluation.