GME-VARCO-VISION-Embedding

About GME-VARCO-VISION-Embedding

<div align="center"> <img src="./varco-vision-Embedding.png" width="100%" style="background-color:white; padding:10px;"/> </div>

GME-VARCO-VISION-Embedding is a multimodal embedding model that computes semantic similarity between text, images, and videos in a high-dimensional embedding space. In particular, the model focuses on video retrieval, which demands greater complexity and contextual understanding compared to image retrieval. It achieves high retrieval accuracy and strong generalization performance across various scenarios, such as scene-based search, description-based search, and question-answering-based search.

Demo Video

Check out our demo videos showcasing our multimodal embedding model in action:

• English Demo Video
• Korean Demo Video

The demo demonstrates how our embedding model works together with an AI agent to search for relevant videos based on user queries and generate responses using the retrieved video content.

Model Architecture and Training Method

GME-VARCO-VISION-Embedding is based on Qwen/Qwen2-VL-7B-Instruct, and uses the parameters of Alibaba-NLP/gme-Qwen2-VL-7B-Instruct to improve the model's general retrieval ability.

1. Fine-tuning (Contrastive Learning) on video preference dataset

To efficiently fine-tune the model, we utilize ShareGPTVideo’s 17𝑘 video preference dataset, which includes prompts, videos, gold answers, and chosen-rejected text pairs. We treat the prompts and videos as queries, and the rejected responses as hard-negatives for the gold answers. . Each query is trained with in-batch negatives as well as one hard negative using the InfoNCE loss. The model is fully fine-tuned for two epochs on 8 A100 GPUs with a batch size of 8, requiring only a few hours for training.

2. Adding Retrieval Vector

To compensate for the insufficiency of training instances and enhance the generalization ability of the fine-tuned model, we compute a retrieval vector 𝜏 by subtracting the weights of the original Qwen/Qwen2-VL-7B-Instruct model from those of Alibaba-NLP/gme-Qwen2-VL-7B-Instruct, a Qwen2-VL based image-text retrieval model. This approach is inspired by Chat Vector, which is a method to equip pre-trained language models with chat capabilities in new languages by adding a vector obtained from the weight difference between a base model and its chat-optimized counterpart.

Performance

Our model achieves state-of-the-art (SOTA) zero-shot performance on the MultiVENT2.0 dataset as of July 2025. See the official leaderboard for detailed results.

<br>

Code Examples

GME-VARCO-VISION-Embedding adopts the inference pipeline of Qwen/Qwen2-VL-7B-Instruct.

Image-Text Retrieval

import torch
import requests
from PIL import Image
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info

model_name = "NCSOFT/GME-VARCO-VISION-Embedding"

model = Qwen2VLForConditionalGeneration.from_pretrained(
    model_name, 
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map="auto",
)

processor = AutoProcessor.from_pretrained(model_name)
tokenizer = processor.tokenizer
device = model.device


qry_msg = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Find a photo of a cat."},
        ],
    },
]

qry_txt = processor.apply_chat_template(
    qry_msg, tokenize=False, add_generation_prompt=True
) + tokenizer.eos_token

qry_input = processor(
            text=[qry_txt],
            padding=True,
            return_tensors="pt",
        ).to(device)  


img_msg = [
    {
        "role": "user",
        "content": [{
            "type": "image",
            "image": "image"
        }]
    }
]

img_txt = processor.apply_chat_template(
    img_msg, tokenize=False, add_generation_prompt=True
) + tokenizer.eos_token


candidate_imgs= [
        # Photo of two cats
        {
            "role": "user",
            "content": [{
                "type": "image",
                "image": "http://images.cocodataset.org/val2017/000000039769.jpg"}]
        },
        # Photo of two dogs
        {
            "role": "user",
            "content": [{
                "type": "image",
                "image": "https://farm1.staticflickr.com/116/290755713_a5de6c1079_z.jpg"}]
        },
        # photo of two people playing baseball
        {
            "role": "user",
            "content": [{
                "type": "image",
                "image": "http://farm3.staticflickr.com/2418/2193688811_d9f5e23bbd_z.jpg"}]
        },
        # Photo of a large crowd in a busy city street
        {
            "role": "user",
            "content": [{
                "type": "image",
                "image":"http://farm7.staticflickr.com/6049/6329686751_997c68fff9_z.jpg"}]
        },
    ]

candidate_images, _ = process_vision_info(candidate_imgs)

image_inputs = processor(
        text=[img_txt] * len(candidate_images),
        images=candidate_images,
        # videos=,
        padding=True,
        return_tensors="pt",
    ).to(device)

with torch.inference_mode():
    qry_emb = model(
        **qry_input, output_hidden_states=True, return_dict=True
    ).hidden_states[-1][:, -1, :]

    img_emb = model(
        **image_inputs, output_hidden_states=True, return_dict=True
    ).hidden_states[-1][:, -1, :]

qry_emb = F.normalize(qry_emb, dim=-1)    
img_emb = F.normalize(img_emb, dim=-1)

score = qry_emb @ img_emb.t()
# tensor([[0.3066, 0.1108, 0.1226, 0.1245]], device='cuda:0', dtype=torch.bfloat16)
# corresponding to the score of photos (cat, dog, baseball, crowd) 

<br>

Video Embedding

vid_message = [
        {
            "role": "user",
            "content": [{
                "type": "video",
                "video": video_path,
                "max_pixels": 262144,
                "fps": 2.0,}]
        }
    ]

video_text = processor.apply_chat_template(
            vid_message, tokenize=False, add_generation_prompt=True
        ) + tokenizer.eos_token

image_input, video_input = process_vision_info(vid_message)

video_input = processor(
        text=[video_text],
        images=image_input,
        videos=video_input,
        padding=True,
        return_tensors="pt",
    ).to(device)

with torch.inference_mode():
    video_emb = model(
        **video_input, output_hidden_states=True, return_dict=True
    ).hidden_states[-1][:, -1, :]

video_emb = F.normalize(video_emb, dim=-1)

<br>

Citation

@misc{park2026vagentinteractivevideosearch,
      title={V-Agent: An Interactive Video Search System Using Vision-Language Models}, 
      author={SunYoung Park and Jong-Hyeon Lee and Youngjune Kim and Daegyu Sung and Younghyun Yu and Young-rok Cha and Jeongho Ju},
      year={2026},
      eprint={2512.16925},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.16925}, 
}

---

license: cc-by-nc-4.0