granite-vision-3.3-2b-embedding

Model Summary: Granite-vision-3.3-2b-embedding is an efficient embedding model based on granite-vision-3.3-2b. This model is specifically designed for multimodal document retrieval, enabling queries on documents with tables, charts, infographics, and complex layouts. The model generates ColBERT-style multi-vector representations of pages. By removing the need for OCR-based text extractions, granite-vision-3.3-2b-embedding can help simplify and accelerate RAG pipelines.

Evaluations: We evaluated granite-vision-3.3-2b-embedding alongside other top colBERT style multi-modal embedding models in the 1B-4B parameter range using two benchmark: Vidore2 and Real-MM-RAG-Bench which aim to specifically address complex multimodal document retrieval tasks.

NDCG@5 - ViDoRe V2

NDCG@5 - REAL-MM-RAG

• Release Date: June 11th 2025
• License: Apache 2.0
• Supported Input Format: Currently the model supports English instructions and images (png, jpeg) as input format.

Intended Use: The model is intended to be used in enterprise applications that involve retrieval of visual and text data. In particular, the model is well-suited for multi-modal RAG systems where the knowledge base is composed of complex enterprise documents, such as reports, slides, images, canned doscuments, manuals and more. The model can be used as a standalone retriever, or alongside a text-based retriever.

Usage

pip install -q torch torchvision torchaudio
pip install transformers==4.50

Then run the code:

from io import BytesIO

import requests
import torch
from PIL import Image
from transformers import AutoProcessor, AutoModel
from transformers.utils.import_utils import is_flash_attn_2_available

device = "cuda" if torch.cuda.is_available() else "cpu"
model_name = "ibm-granite/granite-vision-3.3-2b-embedding"
model = AutoModel.from_pretrained(
                      model_name,
                      trust_remote_code=True,
                      torch_dtype=torch.float16,
                      device_map=device,
                      attn_implementation="flash_attention_2" if is_flash_attn_2_available() else None
                      ).eval()
processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)

# ─────────────────────────────────────────────
# Inputs: Image + Text
# ─────────────────────────────────────────────
image_url = "https://huggingface.co/datasets/mishig/sample_images/resolve/main/tiger.jpg"
print("\nFetching image...")
image = Image.open(BytesIO(requests.get(image_url).content)).convert("RGB")

text = "A photo of a tiger"
print(f"Image and text inputs ready.")

# Process both inputs
print("Processing inputs...")
image_inputs = processor.process_images([image])
text_inputs = processor.process_queries([text])

# Move to correct device
image_inputs = {k: v.to(device) for k, v in image_inputs.items()}
text_inputs = {k: v.to(device) for k, v in text_inputs.items()}

# ─────────────────────────────────────────────
# Run Inference
# ─────────────────────────────────────────────
with torch.no_grad():
    print("🔍 Getting image embedding...")
    img_emb = model(**image_inputs)

    print("✍️ Getting text embedding...")
    txt_emb = model(**text_inputs)

# ─────────────────────────────────────────────
# Score the similarity
# ─────────────────────────────────────────────
print("Scoring similarity...")
similarity = processor.score(txt_emb, img_emb, batch_size=1, device=device)

print("\n" + "=" * 50)
print(f"📊 Similarity between image and text: {similarity.item():.4f}")
print("=" * 50)

Use granite-vision-embedding-3.3-2b for MM RAG

For an example of MM-RAG using granite-vision-3.3-2b-embedding refer to this notebook.

Model Architecture: The architecture of granite-vision-3.3-2b-embedding follows ColPali(https://arxiv.org/abs/2407.01449) approach and consists of the following components:

(1) Vision-Language model : granite-vision-3.3-2b (https://huggingface.co/ibm-granite/granite-vision-3.3-2b).

(2) Projection layer: linear layer that projects the hidden layer dimension of Vision-Language model to 128 and outputs 729 embedding vectors per image.

The scoring is computed using MaxSim-based late interaction mechanism.

Training Data: Our training data is entirly comprised from DocFM. DocFM is a large-scale comprehensive dataset effort at IBM consisting of 85 million document pages extracted from unique PDF documents sourced from Common Crawl, Wikipedia, and ESG (Environmental, Social, and Governance) reports.

Infrastructure: We train granite-vision-3.3-2b-embedding on IBM’s cognitive computing cluster, which is outfitted with NVIDIA A100 GPUs.

Ethical Considerations and Limitations: The use of Large Vision and Language Models involves risks and ethical considerations people must be aware of, including but not limited to: bias and fairness, misinformation, and autonomous decision-making. Granite-vision-3.3-2b-embedding is not the exception in this regard. Although our alignment processes include safety considerations, the model may in some cases produce inaccurate or biased responses. Regarding ethics, a latent risk associated with all Large Language Models is their malicious utilization. We urge the community to use granite-vision-3.3-2b-embedding with ethical intentions and in a responsible way.

Resources

• 📄 Granite Vision technical report here
• 📄 Real-MM-RAG-Bench paper (ACL 2025) here
• 📄 Vidore 2 paper here
• ⭐️ Learn about the latest updates with Granite: https://www.ibm.com/granite
• 🚀 Get started with tutorials, best practices, and prompt engineering advice: https://www.ibm.com/granite/docs/
• 💡 Learn about the latest Granite learning resources: https://ibm.biz/granite-learning-resources