jinaai/jina-clip-v2

The embedding set trained by Jina AI.

Jina CLIP v2: Multilingual Multimodal Embeddings for Texts and Images

Quick Start

Blog | Azure | AWS SageMaker | Google Cloud Platform | API

Intended Usage & Model Info

jina-clip-v2 is a general-purpose multilingual and multimodal (text & image) embedding model.

Multimodal embeddings enable searching and understanding data across different modalities through a coherent representation. They serve as the backbone of neural information retrieval and multimodal GenAI applications.

Built upon jina-clip-v1 and our recently released jina-embeddings-v3, jina-clip-v2 features several significant improvements:

• Improved Performance: v2 shows a 3% performance improvement over v1 in both text-image and text-text retrieval tasks. Similar to v1, v2's text encoder can serve as an effective multilingual long-context dense retriever. It performs on par with our frontier model jina-embeddings-v3 (currently the best multilingual embeddings under 1B parameters on MTEB).
• Multilingual Support: Using the same backbone as jina-embeddings-v3 for the text tower, jina-clip-v2 supports 89 languages for multilingual-image retrieval, showing up to 4% improvement compared to nllb-clip-large-siglip on multilingual image retrieval tasks.
• Higher Image Resolution: v2 now supports 512x512 input image resolution, a significant increase from v1's 224x224. This higher resolution enables better processing of detailed images, improved feature extraction, and more accurate recognition of fine-grained visual elements.
• Matryoshka Representations: v2 allows users to truncate the output dimensions of both text and image embeddings from 1024 down to 64, reducing storage and processing overhead while maintaining strong performance.

Measuring 0.9B parameters, jina-clip-v2 combines two powerful encoders:

• the text encoder Jina-XLM-RoBERTa (the backbone of jina-embeddings-v3) and
• the vision encoder EVA02-L14 (an efficient vision Transformer developed by BAAI).

FEATURE	TEXT ENCODER	IMAGE ENCODER
Base Model	Jina-XLM-RoBERTa	EVA02-L
Parameters	561M	304M
Input Specification	8,192 tokens (max)	512×512 pixels
Min Output Dimensions	64	64
Max Output Dimensions	1,024	1,024
Layers	24	24
Attention Mechanism	FlashAttention2	xFormers
Pooling Strategy	Mean pooling	CLS pooling
Additional Features	89 languages supported	Patch size 14x14

These encoders are jointly trained to create aligned representations of images and text.

CLIP-like models have established themselves as the backbone for general-purpose multimodal applications. With jina-clip-v2, we're taking these capabilities to the next level, breaking down language barriers to deliver more accurate cross-modal understanding and retrieval. We're confident this release delivers a promise in making multimodal search and retrieval both more powerful and more accessible to developers worldwide.

Training, Data, Parameters

An updated version of our technical report with details on jina-clip-v2 is coming soon. Stay tuned!

Faster Inference: FA2, XFormers and bf16

On a CUDA enabled torch environment, the model comes in torch.bfloat16 precision by default. It is highly recommended to install FlashAttention and xFormers to make use of their efficient attention mechanism implementations.

Usage

via Jina AI Embedding API

curl https://api.jina.ai/v1/embeddings \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer [JINA_AI_API_TOKEN]" \
  -d @- <<EOFEOF
  {
    "model": "jina-clip-v2",
    "dimensions": 1024,
    "task": "retrieval.query",
    "normalized": true,
    "embedding_type": "float",
    "input": [
        {
            "text": "غروب جميل على الشاطئ"
        },
        {
            "text": "海滩上美丽的日落"
        },
        {
            "text": "A beautiful sunset over the beach"
        },
        {
            "text": "Un beau coucher de soleil sur la plage"
        },
        {
            "text": "Ein wunderschöner Sonnenuntergang am Strand"
        },
        {
            "text": "Ένα όμορφο ηλιοβασίλεμα πάνω από την παραλία"
        },
        {
            "text": "समुद्र तट पर एक खूबसूरत सूर्यास्त"
        },
        {
            "text": "Un bellissimo tramonto sulla spiaggia"
        },
        {
            "text": "浜辺に沈む美しい夕日"
        },
        {
            "text": "해변 위로 아름다운 일몰"
        },
        {
            "image": "https://i.ibb.co/nQNGqL0/beach1.jpg"
        },
        {
            "image": "https://i.ibb.co/r5w8hG8/beach2.jpg"
        }
    ]
  }
EOFEOF

via transformers

# !pip install transformers einops timm pillow
from transformers import AutoModel

# Initialize the model
model = AutoModel.from_pretrained('jinaai/jina-clip-v2', trust_remote_code=True)

# Corpus
sentences = [
    'غروب جميل على الشاطئ', # Arabic
    '海滩上美丽的日落', # Chinese
    'Un beau coucher de soleil sur la plage', # French
    'Ein wunderschöner Sonnenuntergang am Strand', # German
    'Ένα όμορφο ηλιοβασίλεμα πάνω από την παραλία', # Greek
    'समुद्र तट पर एक खूबसूरत सूर्यास्त', # Hindi
    'Un bellissimo tramonto sulla spiaggia', # Italian
    '浜辺に沈む美しい夕日', # Japanese
    '해변 위로 아름다운 일몰', # Korean
]

# Public image URLs or PIL Images
image_urls = ['https://i.ibb.co/nQNGqL0/beach1.jpg', 'https://i.ibb.co/r5w8hG8/beach2.jpg']

# Choose a matryoshka dimension, set to None to get the full 1024-dim vectors
truncate_dim = 512

# Encode text and images
text_embeddings = model.encode_text(sentences, truncate_dim=truncate_dim)
image_embeddings = model.encode_image(
    image_urls, truncate_dim=truncate_dim
)  # also accepts PIL.Image.Image, local filenames, dataURI

# Encode query text
query = 'beautiful sunset over the beach' # English
query_embeddings = model.encode_text(
    query, task='retrieval.query', truncate_dim=truncate_dim
)

# Text to Image
print('En -> Img: ' + str(query_embeddings @ image_embeddings[0].T))
# Image to Image
print('Img -> Img: ' + str(image_embeddings[0] @ image_embeddings[1].T))
# Text to Text
print('En -> Ar: ' + str(query_embeddings @ text_embeddings[0].T))
print('En -> Zh: ' + str(query_embeddings @ text_embeddings[1].T))
print('En -> Fr: ' + str(query_embeddings @ text_embeddings[2].T))
print('En -> De: ' + str(query_embeddings @ text_embeddings[3].T))
print('En -> Gr: ' + str(query_embeddings @ text_embeddings[4].T))
print('En -> Hi: ' + str(query_embeddings @ text_embeddings[5].T))
print('En -> It: ' + str(query_embeddings @ text_embeddings[6].T))
print('En -> Jp: ' + str(query_embeddings @ text_embeddings[7].T))
print('En -> Ko: ' + str(query_embeddings @ text_embeddings[8].T))

via sentence-transformers

# !pip install sentence-transformers einops timm pillow
from sentence_transformers import SentenceTransformer

# Choose a matryoshka dimension
truncate_dim = 512

# Initialize the model
model = SentenceTransformer(
    'jinaai/jina-clip-v2', trust_remote_code=True, truncate_dim=truncate_dim
)

# Corpus
sentences = [
    'غروب جميل على الشاطئ', # Arabic
    '海滩上美丽的日落', # Chinese
    'Un beau coucher de soleil sur la plage', # French
    'Ein wunderschöner Sonnenuntergang am Strand', # German
    'Ένα όμορφο ηλιοβασίλεμα πάνω από την παραλία', # Greek
    'समुद्र तट पर एक खूबसूरत सूर्यास्त', # Hindi
    'Un bellissimo tramonto sulla spiaggia', # Italian
    '浜辺に沈む美しい夕日', # Japanese
    '해변 위로 아름다운 일몰', # Korean
]

# Public image URLs or PIL Images
image_urls = ['https://i.ibb.co/nQNGqL0/beach1.jpg', 'https://i.ibb.co/r5w8hG8/beach2.jpg']

# Encode text and images
text_embeddings = model.encode(sentences, normalize_embeddings=True)
image_embeddings = model.encode(
    image_urls, normalize_embeddings=True
)  # also accepts PIL.Image.Image, local filenames, dataURI

# Encode query text
query = 'beautiful sunset over the beach' # English
query_embeddings = model.encode(
    query, prompt_name='retrieval.query', normalize_embeddings=True
)  

via the ONNX Runtime

# !pip install transformers onnxruntime pillow
import onnxruntime as ort
from transformers import AutoImageProcessor, AutoTokenizer

# Load tokenizer and image processor using transformers
tokenizer = AutoTokenizer.from_pretrained('jinaai/jina-clip-v2', trust_remote_code=True)
image_processor = AutoImageProcessor.from_pretrained(
    'jinaai/jina-clip-v2', trust_remote_code=True
)

# Corpus
sentences = [
    'غروب جميل على الشاطئ', # Arabic
    '海滩上美丽的日落', # Chinese
    'Un beau coucher de soleil sur la plage', # French
    'Ein wunderschöner Sonnenuntergang am Strand', # German
    'Ένα όμορφο ηλιοβασίλεμα πάνω από την παραλία', # Greek
    'समुद्र तट पर एक खूबसूरत सूर्यास्त', # Hindi
    'Un bellissimo tramonto sulla spiaggia', # Italian
    '浜辺に沈む美しい夕日', # Japanese
    '해변 위로 아름다운 일몰', # Korean
]

# Public image URLs or PIL Images
image_urls = ['https://i.ibb.co/nQNGqL0/beach1.jpg', 'https://i.ibb.co/r5w8hG8/beach2.jpg']

# Tokenize input texts and transform input images
input_ids = tokenizer(sentences, return_tensors='np')['input_ids']
pixel_values = image_processor(image_urls)['pixel_values']

# Start an ONNX Runtime Session
session = ort.InferenceSession('jina-clip-v2/onnx/model.onnx')

# Run inference
output = session.run(None, {'input_ids': input_ids, 'pixel_values': pixel_values})

# Keep the normalised embeddings, first 2 outputs are un-normalized
_, _, text_embeddings, image_embeddings = output

License

This model is licensed to download and run under CC BY-NC 4.0. It is available for commercial use via the Jina Embeddings API, AWS, Azure, and GCP. To download for commercial use, please contact us.

Contact

Join our Discord community and chat with other community members about ideas.

Citation

If you find jina-clip-v2 useful in your research, please cite the following paper:

@misc{2405.20204,
    Author = {Andreas Koukounas and Georgios Mastrapas and Michael Günther and Bo Wang and Scott Martens and Isabelle Mohr and Saba Sturua and Mohammad Kalim Akram and Joan Fontanals Martínez and Saahil Ognawala and Susana Guzman and Maximilian Werk and Nan Wang and Han Xiao},
    Title = {Jina CLIP: Your CLIP Model Is Also Your Text Retriever},
    Year = {2024},
    Eprint = {arXiv:2405.20204},
}