V2PE: Improving Multimodal Long-Context Capability of Vision-Language Models with Variable Visual Position Encoding

April 4, 2026 · View on GitHub

The official implementation of the paper "V2PE: Improving Multimodal Long-Context Capability of Vision-Language Models with Variable Visual Position Encoding".

[🆕 Blog] [📜 Paper] [🤗 HF Models] [📖 HF Datasets]

📖 Summary

The main contributions of this work are as follows:

We construct mixed datasets for VLMs' long-context training and evaluation by augmenting existing multimodal instruction tuning datasets and conduct a thorough investigation into why current VLMs struggle with long-context multimodal inputs, revealing that directly applying LLM positional encoding to visual tokens is ineffective.
We propose Variable Visual Position Encoding (V2PE), a novel positional encoding strategy that employs variable and smaller increments for visual tokens, significantly enhancing VLMs' ability to understand and reason over long multimodal contexts.
We apply our V2PE method and extend training data on the open-source VLM, InternVL2-2B. The fine-tuned VLM performs exceptionally well on both general multimodal benchmarks and long-context multimodal tasks, with the capacity to handle sequences of up to 1M tokens.

🛠️ Installation

See INSTALLATION.md

In addition, using this codebase requires executing the following steps:

Install other requirements:

pip install --upgrade pip  # enable PEP 660 support
pip install -e .

📦 Model Preparation

Our models are built from InternVL2-2B. Please download the above model weights and place them in the pretrained/ folder.

model name	type	download	size
InternVL2-2B	VLM	🤗 HF link	4.4 GB

cd pretrained/
# pip install -U huggingface_hub
huggingface-cli download --resume-download --local-dir-use-symlinks False OpenGVLab/InternVL2-2B --local-dir InternVL2-2B

The directory structure is:

pretrained
└── InternVL2-2B/

🔥 Supervised Fine-tuning

Prepare Training Datasets

Download training and validation dataset from HuggingFace
Organize the data as follows in dataset/:

`dataset/` Directory Structure

dataset
├── annotation
│   ├── long_mr_128k/
│   ├── long_mr_256k/
│   ├── long_mr_32k/
│   ├── long_vqa_32k/
│   ├── milebench_16k/
│   └── milebench_nh/
├── image
│   ├── long_mr
│   │   ├── train/
│   │   └── val/
│   ├── long_vqa
│   │   ├── image
│   │   │   ├── deepform
│   │   │   │   ├── train/
│   │   │   │   └── val/
│   │   │   ├── docvqa
│   │   │   │   ├── train/
│   │   │   │   └── val/
│   │   │   ├── infovqa
│   │   │   │   ├── train/
│   │   │   │   └── val/
│   │   │   ├── kleistercharity
│   │   │   │   ├── train/
│   │   │   │   └── val/
│   │   │   ├── svqa
│   │   │   │   ├── train/
│   │   │   │   └── val/
│   │   │   └── visualmrc
│   │   │       ├── train/
│   │   │       └── val/
│   │   └── paste
│   │       ├── chartqa
│   │       │   ├── train/
│   │       │   └── val/
│   │       ├── clevr
│   │       │   ├── train/
│   │       │   └── val/
│   │       ├── dvqa
│   │       │   ├── train/
│   │       │   └── val/
│   │       ├── gqa
│   │       │   ├── train/
│   │       │   └── val/
│   │       ├── ocrvqa
│   │       │   ├── train/
│   │       │   └── val/
│   │       ├── okvqa
│   │       │   ├── train/
│   │       │   └── val/
│   │       ├── tabfact
│   │       │   ├── train/
│   │       │   └── val/
│   │       ├── textcaps
│   │       │   ├── train/
│   │       │   └── val/
│   │       ├── textvqa
│   │       │   ├── train/
│   │       │   └── val/
│   │       ├── vizwiz
│   │       │   ├── train/
│   │       │   └── val/
│   │       └── wikitablequestions
│   │           ├── train/
│   │           └── val/
│   └── milebench
│       ├── clevr
│       │   └── train/
│       ├── gpr
│       │   └── train/
│       ├── iedit
│       │   └── train/
│       ├── mmcoqa
│       │   └── train/
│       ├── mmqa
│       │   └── train/
│       ├── nh
│       │   └── train/
│       ├── objintercn
│       │   └── train/
│       ├── ocrvqa
│       │   └── train/
│       ├── percept
│       │   └── train/
│       ├── slidevqa
│       │   └── train/
│       ├── spotdiff
│       │   └── train/
│       ├── sta_charades
│       │   └── train/
│       ├── star
│       │   └── train/
│       ├── tqa
│       │   └── train/
│       └── webqa
│           └── train/
└── val
    ├── long_mr_128k/
    ├── long_mr_1m/
    ├── long_mr_256k/
    ├── long_mr_512k/
    ├── long_vqa_32k/
    ├── long_vqa_40k/
    ├── long_vqa_48k/
    ├── long_vqa_56k/
    └── long_vqa_64k/

Start Training

We provide slurm scripts for multi-node multi-GPU training. You can use 32 GPUs to train this model, and it will take approximately 48 hours.

# using 32 GPUs
PARTITION='your partition' GPUS=32 sh shell/internlm2_2b/internvl_chat_v2_internlm2_2b_dynamic_res_v2pe_32k.sh

Training using ring-attention

When training on 256k length or longer dataset, you may need using ring attention to limit GPU memory usage. To use ring attention, you need to set two variables in the training script:

  --chunk_num 8 \
  --attn_type 'ring' \

Here, chunk_num specifies the number of chunks each sample is split into, which are distributed across chunk_num GPUs. The use_chunkTrainer flag indicates that ring attention is used during training.

We provide an example training script that utilizes ring attention at: shell/internlm2_2b/internvl_chat_v2_internlm2_2b_dynamic_res_v2pe_256k.sh. You can run this script with the following command:

# using 32 GPUs
PARTITION='your partition' GPUS=32 sh shell/internlm2_2b/internvl_chat_v2_internlm2_2b_dynamic_res_v2pe_256k.sh

Model	#Param	ChartQA	DocVQA	AI2D	InfoVQA	SQA	POPE	MMMU_val	MMBench_EN	SEED_I	Avg
InternVL2-2B	2.0B	71.7	86.9	74.1	58.9	94.1	85.2	36.3	73.4	70.9	72.4
DeepSeek-VL-1.3B	2.0B	47.4	-	51.5	-	68.4	85.9	33.8	66.4	66.0	-
Qwen2-VL-2B	2.0B	73.5	90.1	74.7	65.5	-	-	41.1	74.9	-	-
Aquila-VL-2B	2.2B	32.0	85.0	75.1	58.3	95.1	83.1	46.9	79.0	73.9	69.8
MiniCPM-V-2	2.8B	55.6	71.9	62.9	-	80.7	86.3	38.2	64.1	67.1	-
Vintern-3B-beta	3.7B	68.3	-	69.1	-	75.0	87.4	46.7	70.6	70.0	-
Llama 3.2 11B	11B	83.4	88.4	91.1	-	-	-	50.7	68.0	-	-
Qwen2-VL-72B	73B	88.3	96.5	88.1	84.5	91.2	87.2	64.5	86.9	77.9	85.0
GPT-4o	-	85.7	92.8	84.7	-	90.1	97.2	69.1	82.1	76.7	-
InternVL2-V2PE-32K	2.0B	76.4	83.9	73.2	55.9	94.9	88.8	36.6	73.5	71.2	72.5

Long-Context MLLM Benchmarks

Model	#Param	MM-NIAH/Image	MM-NIAH/Text	MM-NIAH/Avg	Milebench/T	Milebench/S	Milebench/NI	Milebench/Avg	VideoMME	MVBench
InternVL2-2B	2.0B	23.0	18.9	21.0	58.2	54.5	37.0	49.9	-	-
Phi-3-Vision	2.7B	-	-	-	46.9	50.0	-	-	-	-
OmChat	3.9B	-	-	-	51.4	52.0	-	-	45.9	50.2
LongLLaVA	9B	-	-	-	47.3	46.8	-	-	43.7	49.1
LongLLaVA	13B	-	-	-	52.7	52.1	-	-	51.6	54.6
VILA	13B	14.5	40.5	27.5	-	-	-	-	-	-
Gemini-1.5	-	28.5	82.1	55.2	50.2	58.3	97.9	68.8	69.6	-
GPT-4V	-	-	84.1	-	45.6	58.9	99.4	68.0	59.9	43.5
GPT-4o	-	-	-	-	56.2	63.5	-	-	64.7	-
Claude3-Opus	-	-	-	-	37.4	48.1	85.3	56.9	59.7	-
InternVL2-V2PE-32K	2.0B	78.1	85.7	81.8	65.5	56.4	97.2	72.5	50.7	65.6

mkdir -p data/chartqa && cd data/chartqa

# download images from https://drive.google.com/file/d/1Lm_w6zeET1Hyl_9ks6w5nEsgpoyPHalV/view

# download converted files
wget https://ofasys-wlcb.oss-cn-wulanchabu.aliyuncs.com/Qwen-VL/evaluation/chartqa/train_human.jsonl
wget https://ofasys-wlcb.oss-cn-wulanchabu.aliyuncs.com/Qwen-VL/evaluation/chartqa/train_augmented.jsonl
wget https://ofasys-wlcb.oss-cn-wulanchabu.aliyuncs.com/Qwen-VL/evaluation/chartqa/test_human.jsonl
wget https://ofasys-wlcb.oss-cn-wulanchabu.aliyuncs.com/Qwen-VL/evaluation/chartqa/test_augmented.jsonl

cd ../..

DocVQA val & test

Data Preparation

mkdir -p data/docvqa && cd data/docvqa

# download images and annotations
wget https://datasets.cvc.uab.es/rrc/DocVQA/train.tar.gz --no-check-certificate # (optional)
wget https://datasets.cvc.uab.es/rrc/DocVQA/val.tar.gz --no-check-certificate
wget https://datasets.cvc.uab.es/rrc/DocVQA/test.tar.gz --no-check-certificate

# unzip files
tar -zxvf train.tar.gz
tar -zxvf val.tar.gz
tar -zxvf test.tar.gz

# download converted jsonl files
wget https://ofasys-wlcb.oss-cn-wulanchabu.aliyuncs.com/Qwen-VL/evaluation/docvqa/train.jsonl
wget https://ofasys-wlcb.oss-cn-wulanchabu.aliyuncs.com/Qwen-VL/evaluation/docvqa/val.jsonl
wget https://ofasys-wlcb.oss-cn-wulanchabu.aliyuncs.com/Qwen-VL/evaluation/docvqa/test.jsonl
cd ../..

AI2D test

Data Preparation

mkdir -p data/ai2diagram && cd data/ai2diagram
# download converted files
wget https://huggingface.co/OpenGVLab/InternVL/raw/main/ai2d_test_vlmevalkit.jsonl -O test_vlmevalkit.jsonl
wget https://huggingface.co/OpenGVLab/InternVL/resolve/main/AI2D_TEST.zip && unzip AI2D_TEST.zip

# download images from Google drive (optional, provided by InternLM-XComposer)
# https://drive.google.com/file/d/1dqqa3MnrxMXaU_K9JA6C83je32ibwdOY/view?usp=sharing
# images should be placed in `data/ai2diagram/ai2d/abc_images` and `data/ai2diagram/ai2d/images`
cd ../..

InfoVQA

Data Preparation

Please refer to https://rrc.cvc.uab.es/?ch=17 for details

ScienceQA test

Data Preparation

mkdir -p data/scienceqa/images && cd data/scienceqa/images

# download images
wget https://scienceqa.s3.us-west-1.amazonaws.com/images/test.zip && unzip test.zip

cd ..

# download original questions
wget https://github.com/lupantech/ScienceQA/blob/main/data/scienceqa/problems.json

# download converted files
wget https://ofasys-wlcb.oss-cn-wulanchabu.aliyuncs.com/Qwen-VL/evaluation/scienceqa/scienceqa_test_img.jsonl

cd ../..

POPE

Data Preparation

mkdir -p data/pope && cd data/pope

# make sure you have downloaded COCO images
ln -s ../coco/val2014 ./
wget https://github.com/OpenGVLab/InternVL/releases/download/data/llava_pope_test.jsonl

# download `coco` from POPE
mkdir -p coco && cd coco
wget https://github.com/AoiDragon/POPE/raw/e3e39262c85a6a83f26cf5094022a782cb0df58d/output/coco/coco_pope_adversarial.json
wget https://github.com/AoiDragon/POPE/raw/e3e39262c85a6a83f26cf5094022a782cb0df58d/output/coco/coco_pope_popular.json
wget https://github.com/AoiDragon/POPE/raw/e3e39262c85a6a83f26cf5094022a782cb0df58d/output/coco/coco_pope_random.json
cd ../../..

MMMU

Data Preparation

The evaluation code will automatically download the dataset from huggingface.

MMBench dev & test

Data Preparation

mkdir -p data/mmbench && cd data/mmbench

# download csv files of mmbench
wget http://opencompass.openxlab.space/utils/MMBench/CCBench_legacy.tsv
wget https://download.openmmlab.com/mmclassification/datasets/mmbench/mmbench_dev_20230712.tsv
wget https://download.openmmlab.com/mmclassification/datasets/mmbench/mmbench_dev_cn_20231003.tsv
wget https://download.openmmlab.com/mmclassification/datasets/mmbench/mmbench_dev_en_20231003.tsv
wget https://download.openmmlab.com/mmclassification/datasets/mmbench/mmbench_test_cn_20231003.tsv
wget https://download.openmmlab.com/mmclassification/datasets/mmbench/mmbench_test_en_20231003.tsv

cd ../..

SEED

Data Preparation

mkdir -p data/SEED && cd data/SEED
# 1. Follow the official instructions [Data Preparation for SEED-Bench-1](https://github.com/AILab-CVC/SEED-Bench/blob/main/DATASET.md#data-preparation-for-seed-bench-1)
#    to download the images and the videos. Put images under `./data/SEED/SEED-Bench-image`.
# 2. Extract the video frame in the middle from the downloaded videos, and put them under `./data/SEED/SEED-Bench-image`.
#    LLaVA provided the script [`extract_video_frames.py`](../internvl_chat/tools/extract_video_frames.py) modified from the official one.

wget https://huggingface.co/OpenGVLab/InternVL/raw/main/seed.jsonl
cd ../..

Preparing Long-Context MLLM Benchmarks

MM-NIAH

Data Preparation

Download MM-NIAH dataset from HuggingFace and put the files in dataset/benchmark/MM-NIAH folder.

Unzip images using the following command

tar -xzvf dataset/benchmark/MM-NIAH/mm_niah_test/images.tar.gz -C dataset/benchmark/MM-NIAH/mm_niah_test/
tar -xzvf dataset/benchmark/MM-NIAH/mm_niah_val/annotations.tar.gz -C dataset/benchmark/MM-NIAH/mm_niah_val/

The directory structure should look like this:

dataset
└── benchmark
    └── MM-NIAH
        ├── mm_niah_test
        │   ├── annotations/
        │   └── images/
        └── mm_niah_val/
            ├── annotations/
            └── images/

Milebench

Data Preparation

Download milebench dataset from hugging face

Unzip them using the following command

for file in MileBench_part*.tar.gz
do
tar -xzvf "$file"
done

Put the unzipped files in dataset/benchmark/MileBench folder. The directory structure should look like this:

dataset
└── benchmark
    └── MileBench
        ├── ActionLocalization
        │   ├── images/
        │   └── ActionLocalization.json
        ├── ActionPrediction
        │   ├── images/
        │   └── ActionPrediction.json 
        |── ActionSequence
        │   ...

Evaluation Steps

Evaluating General MLLM Benchmarks

Evaluation

For all general MLLM benchmarks, you can only run this one scripts to get all results.

# use STRIDE=64 as an example
STRIDE=64 sh scripts/evaluate_auto.sh <checkpoint> --rope_pos_id_version v2pe_fix --rope_pos_id_stride 64

Evaluating Long-Context MLLM Benchmarks

Evaluation for milebench

# use STRIDE=64 as an example
STRIDE=64 sh scripts/evaluate_milebench.sh <checkpoint> --rope_pos_id_version v2pe_fix --rope_pos_id_stride 64

Evaluation for mm_niah

# use STRIDE=64 as an example
STRIDE=64 sh scripts/evaluate_mmniah.sh <checkpoint> --rope_pos_id_version v2pe_fix --rope_pos_id_stride 64

Evaluation for mm_niah-1M

# use STRIDE=64 as an example
STRIDE=64 sh scripts/evaluate_mmniah_long.sh <checkpoint> --rope_pos_id_version v2pe_fix --rope_pos_id_stride 64

Evaluation for long-vqa

# use STRIDE=64 as an example
STRIDE=64 GROUP=32 GPUS_PER_TASK=1 sh scripts/evaluate_longvqa.sh <checkpoint> --rope_pos_id_version v2pe_fix --rope_pos_id_stride 64
STRIDE=64 GROUP=40 GPUS_PER_TASK=2 sh scripts/evaluate_longvqa.sh <checkpoint> --rope_pos_id_version v2pe_fix --rope_pos_id_stride 64
STRIDE=64 GROUP=48 GPUS_PER_TASK=2 sh scripts/evaluate_longvqa.sh <checkpoint> --rope_pos_id_version v2pe_fix --rope_pos_id_stride 64
STRIDE=64 GROUP=56 GPUS_PER_TASK=4 sh scripts/evaluate_longvqa.sh <checkpoint> --rope_pos_id_version v2pe_fix --rope_pos_id_stride 64
STRIDE=64 GROUP=64 GPUS_PER_TASK=4 sh scripts/evaluate_longvqa.sh <checkpoint> --rope_pos_id_version v2pe_fix --rope_pos_id_stride 64

🎫 License

This project is released under the MIT License.

🖊️ Citation

If you find this work helpful in your research, please consider citing:

@misc{ge2024v2peimprovingmultimodallongcontext,
      title={V2PE: Improving Multimodal Long-Context Capability of Vision-Language Models with Variable Visual Position Encoding}, 
      author={Junqi Ge and Ziyi Chen and Jintao Lin and Jinguo Zhu and Xihui Liu and Jifeng Dai and Xizhou Zhu},
      year={2024},
      eprint={2412.09616},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2412.09616}, 
}

V2PE: Improving Multimodal Long-Context Capability of Vision-Language Models with Variable Visual Position Encoding

📖 Summary

🛠️ Installation

📦 Model Preparation

🔥 Supervised Fine-tuning

Prepare Training Datasets

Start Training

Training using ring-attention

📊 Evaluation

Evaluation results in paper

Evaluation results of our released model

❓ How to Evaluate

Preparing General MLLM Benchmarks

ChartQA test-human & test-augmented

DocVQA val & test

AI2D test

InfoVQA

ScienceQA test

POPE

MMMU

MMBench dev & test

SEED

Preparing Long-Context MLLM Benchmarks

MM-NIAH

Milebench

Evaluation Steps

Evaluating General MLLM Benchmarks

Evaluating Long-Context MLLM Benchmarks

🎫 License

🖊️ Citation