File size: 7,007 Bytes
7fb29cc ed49a05 7fb29cc ed49a05 019fd32 ed49a05 019fd32 ed49a05 afd746d ed49a05 019fd32 ed49a05 019fd32 ed49a05 019fd32 ed49a05 afd746d ed49a05 019fd32 ed49a05 019fd32 ed49a05 019fd32 ed49a05 afd746d ed49a05 019fd32 ed49a05 3ba046a ed49a05 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 |
---
license: apache-2.0
language:
- en
metrics:
- recall
base_model:
- friedrichor/Unite-Base-Qwen2-VL-7B
tags:
- sentence-transformers
- sentence-similarity
- transformers
- multimodal
- retrieval
- feature-extraction
- image-text-to-text
- video-text-to-text
- any-to-any
datasets:
- friedrichor/Unite-Instruct-Retrieval-Train
---
## Modality Curation: Building Universal Embeddings for Advanced Multimodal Information Retrieval
[](https://opensource.org/licenses/Apache-2.0)
[](https://huggingface.co/papers/2505.19650)
[](https://github.com/friedrichor/UNITE)
[](https://friedrichor.github.io/projects/UNITE)
[](https://huggingface.co/collections/friedrichor/unite-682da30c4540abccd3da3a6b)
<p align="center">
<img src="https://raw.githubusercontent.com/friedrichor/UNITE/main/assets/overall_task.png" alt="task" width="80%">
</p>
## UNITE: UNIversal mulTimodal Embedder
<p align="center">
<img src="https://raw.githubusercontent.com/friedrichor/UNITE/main/assets/overall_model.png" alt="model_arch" width="100%">
</p>
**Support Modalities and Tasks:**
⚡ **Unified Multimodal Representations:** *text*, *image*, *video*, and *their fusion*.
⚡ **Enhancements in Diverse Tasks:** *coarse-grained retrieval*, *fine-grained retrieval* (Recommended UNITE-Base), and *instruction-based retrieval* (Recommended UNITE-Instruct)
## Requirements
```bash
pip install torch==2.5.0 torchvision==0.20.0 torchaudio==2.5.0
pip install flash-attn --no-build-isolation
pip install transformers==4.47.1
pip install qwen-vl-utils[decord]==0.0.8
```
## Quickstart
```bash
# get inference code from https://huggingface.co/friedrichor/Unite-Base-Qwen2-VL-2B/tree/main/inference_demo
cd inference_demo
```
### Load Model
```python
import torch
from transformers import AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info
from modeling_unite import UniteQwen2VL
model_path = 'friedrichor/Unite-Instruct-Qwen2-VL-7B'
model = UniteQwen2VL.from_pretrained(
model_path,
device_map="cuda",
torch_dtype=torch.bfloat16,
)
# We recommend enabling flash_attention_2 for better acceleration and memory saving.
# model = UniteQwen2VL.from_pretrained(
# model_path,
# device_map="cuda",
# torch_dtype=torch.bfloat16,
# attn_implementation='flash_attention_2',
# low_cpu_mem_usage=True,
# )
tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
processor = AutoProcessor.from_pretrained(model_path, min_pixels=256*28*28, max_pixels=1280*28*28)
def process_messages(msg):
text = processor.apply_chat_template(msg, tokenize=False, add_generation_prompt=True) + "<|endoftext|>"
image_inputs, video_inputs = process_vision_info(msg)
inputs = processor(
text=[text],
images=image_inputs,
videos=video_inputs,
padding=True,
return_tensors="pt",
)
inputs = inputs.to("cuda")
return inputs
```
### Inference
<details>
<summary>Image-Text Retrieval</summary>
```python
messages_txt = [
{
"role": "user",
"content": [
{"type": "text", "text": "The book titled 'Riding with Reindeer - A Bicycle Odyssey through Finland, Lapland, and the Arctic' provides a detailed account of a journey that explores the regions of Lapland and the Arctic, focusing on the experience of riding with reindeer."},
{"type": "text", "text": "\nSummary above sentence in one word:"},
],
}
]
messages_img = [
{
"role": "user",
"content": [
{"type": "image", "image": "./examples/518L0uDGe0L.jpg"},
{"type": "text", "text": "\nSummary above image in one word:"},
],
}
]
inputs_txt = process_messages(messages_txt)
inputs_img = process_messages(messages_img)
with torch.no_grad():
embeddings_txt = model(**inputs_txt) # [1, 3584]
embeddings_img = model(**inputs_img) # [1, 3584]
print(torch.matmul(embeddings_txt, embeddings_img.T))
# tensor([[0.7578]], dtype=torch.bfloat16)
```
</details>
<details>
<summary>Video-Text Retrieval</summary>
```python
messages_txt = [
{
"role": "user",
"content": [
{"type": "text", "text": "Timelapse of stormy clouds over open sea and snowcapped mountain"},
{"type": "text", "text": "\nSummary above sentence in one word:"},
],
}
]
messages_vid = [
{
"role": "user",
"content": [
{
"type": "video",
"video": "./examples/stock-footage-timelapse-of-stormy-clouds-over-open-sea-and-snowcapped-mountain.mp4",
"max_pixels": 360 * 420,
"fps": 1,
"max_frames": 32
},
{"type": "text", "text": "\nSummary above video in one word:"},
],
}
]
inputs_txt = process_messages(messages_txt)
inputs_vid = process_messages(messages_vid)
with torch.no_grad():
embeddings_txt = model(**inputs_txt) # [1, 3584]
embeddings_vid = model(**inputs_vid) # [1, 3584]
print(torch.matmul(embeddings_txt, embeddings_vid.T))
# tensor([[0.4883]], dtype=torch.bfloat16)
```
</details>
<details>
<summary>Fused-Modal Retrieval</summary>
```python
messages_qry = [
{
"role": "user",
"content": [
{"type": "image", "image": "./examples/oven_05011373.jpg"},
{"type": "text", "text": "What is the name of this place?"},
{"type": "text", "text": "\nSummary above sentence and image in one word:"},
],
}
]
messages_tgt = [
{
"role": "user",
"content": [
{"type": "image", "image": "./examples/Q673659.jpg"},
{"type": "text", "text": "Marina Beach."},
{"type": "text", "text": "\nSummary above sentence and image in one word:"},
],
}
]
inputs_qry = process_messages(messages_qry)
inputs_tgt = process_messages(messages_tgt)
with torch.no_grad():
embeddings_qry = model(**inputs_qry) # [1, 3584]
embeddings_tgt = model(**inputs_tgt) # [1, 3584]
print(torch.matmul(embeddings_qry, embeddings_tgt.T))
# tensor([[0.6719]], dtype=torch.bfloat16)
```
</details>
## Citation
If you find our work helpful, feel free to give us a cite.
```
@article{kong2025modality,
title={Modality Curation: Building Universal Embeddings for Advanced Multimodal Information Retrieval},
author={Kong, Fanheng and Zhang, Jingyuan and Liu, Yahui and Zhang, Hongzhi and Feng, Shi and Yang, Xiaocui and Wang, Daling and Tian, Yu and W., Victoria and Zhang, Fuzheng and Zhou, Guorui},
journal={arXiv preprint arXiv:2505.19650},
year={2025}
}
``` |