---
title: VisME-Qwen25VL-7B
canonical_url: "https://www.modelscope.cn/models/Christin/VisME-Qwen25VL-7B"
md_url: "https://www.modelscope.cn/models/Christin/VisME-Qwen25VL-7B.md"
repository: Christin/VisME-Qwen25VL-7B
last_updated: 2026-08-01
license: apache-2.0
pipeline_tag: image-text-to-text
tasks:
  - image-text-to-text
model_type:
  - qwen2_5_vl
architectures:
  - Qwen2_5_VLForConditionalGeneration
base_model:
  - Qwen/Qwen2.5-VL-7B-Instruct
base_model_relation: finetune
parameters: 7.7B
tensor_type:
  - BF16
library_name:
  - pytorch
  - safetensors
downloads: 17
stars: 1
tags:
  - embedding
  - multimodal
  - vision-language
  - feature-extraction
  - mmeb
  - mveb
---

# VisME-Qwen25VL-7B

> VisME-Qwen25VL-7B - Christin 在 ModelScope 开源的模型。VisME (Qwen2.5-VL-7B)

Christin/VisME-Qwen25VL-7B 是 ModelScope 魔搭社区上的 7.7B 参数image-text-to-text模型，采用 apache-2.0 许可，基于 Qwen/Qwen2.5-VL-7B-Instruct 构建。

- **Repository**: Christin/VisME-Qwen25VL-7B
- **License**: apache-2.0
- **Tasks**: image-text-to-text
- **Parameters**: 7.7B
- **Base model**: Qwen/Qwen2.5-VL-7B-Instruct
- **Tags**: embedding, multimodal, vision-language, feature-extraction, mmeb, mveb
- **Downloads**: 17
- **Stars**: 1
- **Last updated**: 2026-08-01

Source: https://www.modelscope.cn/models/Christin/VisME-Qwen25VL-7B

---

# VisME (Qwen2.5-VL-7B)

[English](README-en.md) | 简体中文

**VisME** 是我们 CVPR 2026 论文 [*Illuminating Visual Identity in Universal Multimodal Embeddings*](https://openaccess.thecvf.com/content/CVPR2026/html/Cao_Illuminating_Visual_Identity_in_Universal_Multimodal_Embeddings_CVPR_2026_paper.html) 提出的通用多模态嵌入模型。VisME 基于 [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) 构建，可为图片、文本以及图文对生成稠密嵌入，在通用多模态基准（**MMEB**）和视觉 ID 检索（**MVEB**）上均有出色表现。

<p align="center">
  <img src="assets/teaser.jpg" alt="VisME teaser" width="95%">
</p>

## 亮点

- **ID 感知检索** —— 通过 [MVEB](https://chrisclear3.github.io/MVEB/) 基准（4 类元任务、28 个测试子集）引入视觉 ID 监督进行训练。
- **通用能力强** —— 在 **MMEB-v1** 与 **MVEB** 上联合训练，在标准 MMEB 任务上保持有竞争力的表现。

## 性能

下图为 **MMEB**（Cls / VQA / Ret / Grd）与 **MVEB**（ID-Rec / Re-ID / ID-Grd / ID-Edit）上的结果。指标与论文一致（各任务组内所有数据集的平均 Precision@1）。

<p align="center">
  <img src="assets/performance.jpg" alt="VisME performance on MMEB and MVEB" width="95%">
</p>

## 训练数据

VisME 在以下数据的混合集上训练：

| 基准 | 作用 |
|------|------|
| [**MMEB-v1**](https://github.com/TIGER-AI-Lab/VLM2Vec) | 通用多模态嵌入任务（分类、VQA、检索、定位） |
| [**MVEB**](https://chrisclear3.github.io/MVEB/) | 覆盖识别、Re-ID、定位与编辑的视觉 ID 检索 |

**MVEB** 数据集已开源，训练代码即将发布。

## 使用方法

```python
import torch
import torch.nn.functional as F
from io import BytesIO
from urllib.request import Request, urlopen
from PIL import Image
from visme import VisME

model = VisME("path/to/this/repo")
model = model.cuda().eval()

instruction = "Represent the face with the following text."
text = "Retrieve all images with the same cartoon character."

samples = {
    "SpiderMan_comic_E616": "https://static.wikia.nocookie.net/spiderman/images/a/ad/Peter_Parker_%28Earth-616%29_017.png/revision/latest?cb=20210807043502",
    "SpiderMan_promo_E199999": "https://static.wikia.nocookie.net/marveldatabase/images/2/28/Peter_Parker_%28Earth-199999%29_from_Spider-Man_No_Way_Home_promotional_art_002.jpg/revision/latest/scale-to-width-down/1000?cb=20230730084204",
    "Toxin": "https://static.wikia.nocookie.net/superheroes/images/5/53/Toxin.jpg/revision/latest?cb=20240813171932",
}

def load_image(url: str) -> Image.Image:
    req = Request(url, headers={"User-Agent": "Mozilla/5.0"})
    with urlopen(req, timeout=30) as resp:
        return Image.open(BytesIO(resp.read())).convert("RGB")

names = list(samples.keys())
batch = [
    {"image": load_image(samples[name]), "text": text, "instruction": instruction}
    for name in names
]

with torch.no_grad():
    embeddings = model.encode_input(batch)  # 形状: (3, 3584)，已做 L2 归一化

# 两两余弦相似度
# SpiderMan_comic_E616  <-> SpiderMan_promo_E199999
# SpiderMan_comic_E616  <-> Toxin
# SpiderMan_promo_E199999 <-> Toxin
for i in range(len(names)):
    for j in range(i + 1, len(names)):
        sim = F.cosine_similarity(embeddings[i : i + 1], embeddings[j : j + 1]).item()
        print(f"cosine_similarity({names[i]}, {names[j]}) = {sim:.4f}")
```

## 评测

MMEB 与 MVEB 的基准评测代码**即将发布**。在此之前，完整的实验设置与基线请参见论文。

## 引用

如果 VisME 或 MVEB 对您有帮助，请引用：

```bibtex
@inproceedings{cao2026illuminating,
  title={Illuminating Visual Identity in Universal Multimodal Embeddings},
  author={Cao, Jiawei and Feng, Junyi and Hua, Jiashen and Huang, Ziheng and Deng, Bing and Wu, Kaijie and Gu, Chaochen and Ye, Jieping},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={8737--8748},
  year={2026}
}
```

## 许可

本模型基于 [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) 构建。请遵循基础模型以及训练所用各数据集的许可条款。