Open Visual-Pruning Suite is the official implementation of SemVID and a unified evaluation codebase for training-free visual-token pruning on video temporal grounding and VideoQA. SemVID preserves a compact, temporally connected evidence chain by allocating tokens to query-relevant objects, motion transitions, and contextual anchors.

Headline result. With Qwen3-VL-4B, SemVID retains up to 95.4% of the original mIoU using 12.5% of visual tokens, while delivering up to 5.8× prefill speedup.

Why SemVID?

Video temporal grounding needs exact event boundaries and evidence that remains connected across time. Pruning objectives designed for VideoQA often preserve a few salient frames while dropping boundary cues or intermediate relay tokens. SemVID addresses this mismatch through:

• Evidence Retention (ER): keeps query-critical visual evidence, especially around temporal boundaries.
• Connectivity Strength (CS): preserves cross-frame attention paths for long-range evidence aggregation.
• Training-free inference: changes token selection without updating backbone weights.

Quick start

1. Install

conda create -n openvps python=3.11 -y
conda activate openvps

pip install torch==2.8.0 torchvision==0.23.0 \
  --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt
pip install flash-attn==2.8.3 --no-build-isolation

See Installation for CUDA, FFmpeg, hardware, and troubleshooting notes.

2. Prepare data

Open the matching preset under configs/_base_/datasets and set eval_data_path and eval_video_folder to your annotation and video locations. LongVideoBench also requires subtitle_folder:

# configs/_base_/datasets/charades_sta.py
eval_args = dict(
    eval_data_path="/path/to/Charades/sta_annotation/charades_sta_test.txt",
    eval_video_folder="/path/to/Charades/rgb_videos_30fps_480",
    # Keep the remaining released settings unchanged.
)

Alternatively, keep the portable defaults and mirror this layout under one root:

data/
|-- ActivityNet/
|   |-- captions/val_2.json
|   `-- rgb_videos_15fps_short256/
|-- Charades/
|   |-- sta_annotation/charades_sta_test.txt
|   `-- rgb_videos_30fps_480/
|-- VideoMME/
|   |-- test-00000-of-00001.parquet
|   `-- data/
`-- LongVideoBench/
    |-- lvb_val.json
    |-- videos/
    `-- subtitles/

Then pass the parent directory with --data-root; the launcher rebases annotation, video, and subtitle paths automatically. Official download links, exact layouts, and path rules are in Data preparation.

3. Evaluate

CUDA_VISIBLE_DEVICES=0 python tools/evaluate.py \
  configs/SemVID/qwen3_vl_4b_charades.py \
  --data-root /path/to/data \
  --gpus 1

Results are written to the config's training_args.output_dir. Multi-GPU launch, config overrides, dataset switching, and output formats are documented in Evaluation.

Included methods

Method	Qwen3-VL	Qwen2.5-VL	LLaVA-OneVision	Public config
SemVID	✓	✓	✓	configs/SemVID
Sampling baselines	✓	-	-	configs/Baseline
FastVID	✓	✓	-	configs/FastVID
VisionZip	✓	✓	-	configs/VisionZip
ToME	-	✓	-	configs/ToME
TokenSculpt	-	✓	-	configs/TokenSculpt
DART	-	✓	-	configs/DART
VScan	-	✓	-	configs/VScan
PruneVID	-	✓	-	configs/PruneVID

The method presets preserve the released code settings. Dataset and model paths remain portable and can be changed without editing model source files.

Evidence analysis

The paper's ER and CS implementations live in analysis/evidence_metrics.py and are called directly by the Qwen evaluator. A full-token pass establishes the reference landing distribution; a pruned pass then produces per-sample attn_metrics.json files.

python analysis/evidence_metrics.py analysis_outputs/qwen3_charades/samples

See ER/CS analysis for the complete two-pass workflow and metric definitions.

Repository layout

OpenVPS/
|-- analysis/       # Evidence Retention and Connectivity Strength
|-- configs/        # Method-oriented, composable experiment configs
|-- docs/           # Installation, data, evaluation, and analysis guides
|-- src/            # Datasets, evaluators, and model implementations
|-- tools/
|   `-- evaluate.py # Evaluation entry point
|-- CITATION.cff
|-- LICENSE
`-- requirements.txt

Documentation

Guide	Covers
Installation	Environment, CUDA stack, Transformers constraint, verification
Data preparation	Official datasets, directory structure, path configuration
Evaluation	Config composition, single/multi-GPU runs, overrides, outputs
ER/CS analysis	Full-token reference pass, pruned pass, metric aggregation

Citation

@article{li2026keeping,
  title     = {Keeping the Evidence Chain: Semantic Evidence Allocation for Training-Free Token Pruning in Video Temporal Grounding},
  author    = {Li, Jiaqi and Zheng, Shuntian and Shen, Yixian and Huang, Jia-Hong and Lu, Xiaoman and Ni, Minzhe and Guan, Yu},
  journal   = {arXiv preprint arXiv:2603.05663},
  year      = {2026}
}

Acknowledgements

This project builds on Hugging Face Transformers, Qwen-VL, LLaVA-OneVision, OpenTAD, and the visual-token pruning methods included in the suite.

License

Code is released under the Apache License 2.0. Model weights and datasets remain subject to their original licenses and access terms; see NOTICE.