🤿 DIVE: Dynamic Iterative Visual Evidence Construction
for Efficient Vision-Language Models

Welcome to star🌟 this repo or cite✨ the paper if you find it interesting😊

TL;DR: We introduce DIVE, a token pruning method that iteratively construct visual evidence to identify and remove redundant visual tokens.

✨ Key Highlights

• ⚡ Efficient Inference: Achieves a 1.68× end-to-end speedup while maintaining near-full performance.
• 🎯 Extreme Compression: Retains 98.2% performance with only 64 visual tokens (88.9% reduction).
• 🔄 Dynamic Selection: Selects complementary, prompt-relevant visual evidence through iterative residual updates.
• 🌐 Broad Generalization: Consistently improves image and video understanding across diverse VLM architectures.

🔥 News

• 2026.08.01 🤗🤗 We have released DIVE, a training-free method for efficient visual-token reduction in vision-language models.
• 2026.07.29 🤗🤗 The DIVE framework, LLaVA evaluation scripts, and Qwen2-VL and LLaVA-OneVision integrations are now available.

👀 Overview

DIVE provides a native LLaVA-1.5/1.6 language-model path using the DIVE selector, together with Qwen2-VL and LLaVA-OneVision integrations.

🛠 Preparation

LLaVA

conda create -n DIVE python=3.10 -y
conda activate DIVE
pip install -e .
pip install flash-attn --no-build-isolation

Download the benchmark assets from their official sources and extract them under ./playground/data/eval.

Qwen2-VL

conda create -n DIVE_Qwen2VL python=3.10 -y
conda activate DIVE_Qwen2VL
cd othermodels/qwen2_vl/transformers && pip install -e .
pip install accelerate qwen-vl-utils[decord]
pip install flash-attn --no-build-isolation
cd ../../../lmms-eval && pip install -e .

🎯 Usage

LLaVA-1.5

bash scripts/v1_5/eval/[benchmark].sh [reduction_ratio] [max_num_trunction]

CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/textvqa.sh 0.778 128
CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/pope.sh 0.778 128
CUDA_VISIBLE_DEVICES=0 bash scripts/v1_5/eval/mme.sh 0.778 128

LLaVA-NeXT

bash scripts/v1_6/eval/[benchmark].sh [reduction_ratio] [max_num_trunction]

CUDA_VISIBLE_DEVICES=0 bash scripts/v1_6/eval/pope.sh 0.778 128

Qwen2-VL

cd othermodels/qwen2_vl
bash eval_scripts/lmms_eval.sh True [reduction_ratio]

🔑 License

This project is released under the Apache License 2.0. Original upstream copyright and license notices are retained in the source files.

👍 Acknowledgment

DIVE thanks the LLaVA, QwenVL, DART, and lmms-eval communities for their open-source contributions.