ForensicZip: More Tokens are Better but Not Necessary in Forensic Vision-Language Models

Yingxin Lai¹, Zitong Yu^1⋆, Jun Wang^1⋆, Linlin Shen², Yong Xu³, and Xiaochun Cao⁴

¹ Great Bay University
² Shenzhen University
³ Harbin Institute of Technology
⁴ School of Cyber Science and Technology, Sun Yat-sen University

🔍 Overview

Multimodal Large Language Models (MLLMs) enable interpretable multimedia forensics by generating textual rationales for forgery detection. However, processing dense visual sequences incurs high computational cost, especially for high-resolution images and videos. Existing visual token pruning methods are mostly semantic-driven: they preserve salient objects while often discarding background regions where manipulation traces such as high-frequency anomalies and temporal jitters reside.

To address this issue, we introduce ForensicZip, a training-free framework that reformulates token compression from a forgery-driven perspective. ForensicZip models temporal token evolution as a Birth-Death Optimal Transport problem with a slack dummy node, quantifying physical discontinuities associated with transient generative artifacts. The final forensic score further integrates transport-based novelty with high-frequency priors, allowing forensic evidence to be preserved under large-ratio compression.

On deepfake and AIGC benchmarks, ForensicZip delivers strong detection performance at aggressive compression ratios, achieving 2.97× speedup and over 90% FLOPs reduction at 10% token retention while maintaining state-of-the-art accuracy.

Figure 1. Overview of the ForensicZip framework. The method preserves forgery-relevant evidence under aggressive token compression by combining transport-based novelty with forensic priors.

🧱 Repository Structure

• forensiczip/ — method implementation and helper utilities
• fakevlm/ — FakeVLM-compatible skeleton modules
• scripts/ — evaluation entrypoints
• docs/ — running and data preparation notes
• imgs/ — method figures

🛠️ Installation

conda create -n forensiczip python=3.10 -y
conda activate forensiczip
pip install -r requirements.txt

If you already have a compatible environment, you can reuse it directly.

🚀 Running

1. FakeClue Evaluation

MODEL_PATH_7B=<MODEL_PATH> \
FAKECLUE_TEST_JSON=<FAKECLUE_JSON> \
FAKECLUE_DATA_BASE=<FAKECLUE_MEDIA_DIR> \
CUDA_DEVICES=0 \
PYTHON_BIN=python \
bash scripts/eval_forensiczip_fakeclue.sh

2. LOKI Evaluation

MODEL_PATH_7B=<MODEL_PATH> \
LOKI_JSON_DIR=<LOKI_JSON_DIR> \
LOKI_MEDIA_ROOT=<LOKI_MEDIA_ROOT> \
CUDA_DEVICES=0 \
PYTHON_BIN=python \
bash scripts/eval_forensiczip_loki.sh

3. Common Options

• RETENTION_RATIOS_STR
• VAL_BATCH_SIZE
• WORKERS
• MAX_LENGTH
• MAX_NEW_TOKENS
• FORENSICZIP_SELECT_LAYER
• FORENSICZIP_BIRTH_COST
• FORENSICZIP_DEATH_COST
• FORENSICZIP_SINKHORN_EPS
• FORENSICZIP_SINKHORN_ITERS
• FORENSICZIP_EMA_BETA
• FORENSICZIP_BIRTH_WEIGHT
• FORENSICZIP_POS_LAMBDA
• FORENSICZIP_FORENSIC_ETA

Detailed usage notes are available in docs/running.md.

📦 External Resources

These resources are used by this repository but are not introduced by this work.

• 
  FakeVLM checkpoint used for evaluation.

• 
  FakeClue dataset used in evaluation.

• 
  Upstream framework that provides the base model and evaluation structure.

See docs/data_preparation.md for the expected local file layout.

🙏 Acknowledgement

This codebase is built on top of FakeVLM. We thank the FakeVLM project for providing the base model and evaluation structure used in this release.

📝 Citation

If you find this repository useful, please consider citing:

@article{lai2026forensiczip,
  title={ForensicZip: More Tokens are Better but Not Necessary in Forensic Vision-Language Models},
  author={Lai, Yingxin and Yu, Zitong and Wang, Jun and Shen, Linlin and Xu, Yong and Cao, Xiaochun},
  journal={arXiv preprint},
  year={2026}
}

📬 Contact

For questions about this repository, please contact: yingxinlai2@gmail.com