QViC-MF-7B

This is the official model for the paper "Question-guided Visual Compression with Memory Feedback for Long-Term Video Understanding" (CVPR 2026).

Paper: arXiv
Code: GitHub
Authors: Sosuke Yamao*, Natsuki Miyahara*, Yuankai Qi, Shun Takeuchi
Affiliation: Fujitsu Research / Macquarie University

Model Description

QViC-MF-7B is a LoRA adapter trained on top of LLaVA-Video-7B-Qwen2 for long-term video understanding. It introduces Question-guided Multimodal Selective Attention (QMSA), which learns to preserve visual information related to the given question from both the current clip and past related frames stored in a context memory. The compressor and memory feedback work iteratively for each clip of the entire video, enabling strong performance on tasks that require understanding complete events (e.g., temporal ordering).

How to Use

Please refer to the official code repository for inference and evaluation scripts.

Model Details

Property	Value
Base Model	lmms-lab/LLaVA-Video-7B-Qwen2
Architecture	LlavaQwenForCausalLM
Vision Encoder	SigLIP-SO400M (google/siglip-so400m-patch14-384)
LLM	Qwen2-7B
Adapter	LoRA (r=64, alpha=16)
Context Memory Length	256 tokens
Precision	bfloat16

Citation

@inproceedings{yamao2026qvicmf,
  title={Question-guided Visual Compression with Memory Feedback for Long-Term Video Understanding},
  author={Yamao, Sosuke and Miyahara, Natsuki and Qi, Yuankai and Takeuchi, Shun},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}