ManiVidLens-4B
ManiVidLens unifies forgery detection, artifact grounding, and anomaly explanation for manipulated videos.
This final Stage 2 checkpoint uses Qwen3-VL-4B-Instruct, SAM2 Hiera-Large, and AIDE ProGAN with 5 input frames. The language-model adapters are merged. The repository includes model weights, tokenizer and processor files, runtime model definitions, and the frozen AIDE encoder in aide_pretrain.pth. Training logs, optimizer states, and scheduler states are excluded.
Inference
Install the dependencies in the code repository, then run from its root:
hf download zzzyzh/ManiVidLens-4B --local-dir ckpts/ManiVidLens-4B
bash scripts/infer.sh \
--model ckpts/ManiVidLens-4B \
--input data/val.json \
--video-root data/videos \
--output outputs/results/ManiVidLens-4B \
--gpus 1 \
--embedding-model sentence-transformers/paraphrase-MiniLM-L6-v2
The annotation video paths are relative to --video-root. The command runs detection, explanation, and segmentation. Keep aide_pretrain.pth in the model directory; separate foundation checkpoints are not needed for inference.
The model uses the ManiVidLens codebase and its bundled AIDE and SAM2 implementations. See the code repository for setup and input formatting.
Acknowledgements
This model builds on Qwen3-VL, SAM2, and AIDE.
Citation
@article{kang2026manivid,
title={ManiVid: Unified and Explainable Forensic Analysis of Manipulated Videos},
author={Kang, Hengrui and Yan, Zhonghao and Yang, Yuxuan and Jing, Ruoyan and Guo, Yuncheng and Chen, Hao and Liang, Kongming and Ma, Zhanyu and He, Conghui and Li, Weijia},
journal={arXiv preprint arXiv:2609.30934},
year={2026}
}
- Downloads last month
- -
Model tree for zzzyzh/ManiVidLens-4B
Base model
Qwen/Qwen3-VL-4B-Instruct