vdr-2b-v1

vdr-2b-v1 is an english only embedding model designed for visual document retrieval. It encodes document page screenshots into dense single-vector representations, this will effectively allow to search and query visually rich multilingual documents without the need for any OCR, data extraction pipelines, chunking...

• Trained on the 🇬🇧 English vdr-multi-train subset: extensive training dataset of 100k high-quality english samples.

• Low VRAM and Faster Inference: achieves better results on synthetic Vidore benchmarks with just 30% of the base model image resolution. This results in 3x faster inference and much lower VRAM usage.

• Matryoshka Representation Learning: You can reduce the vectors size 3x and still keep 98% of the embeddings quality.

The multilingual version is available here. To know more about both models, read the announcement blogpost.

Usage

The model uses bf16 tensors and allocates ~4.4GB of VRAM when loaded. You can easily run inference and generate embeddings using 768 image patches and a batch size of 16 even on a cheap NVIDIA T4 GPU. This table reports the memory footprint (GB) under conditions of different batch sizes with HuggingFace Transformers and maximum 768 image patches.

Batch Size	GPU Memory (GB)
4	6.9
8	8.8
16	11.5
32	19.7

You can generate embeddings with this model in many different ways:

via LlamaIndex

pip install -U llama-index-embeddings-huggingface

from llama_index.embeddings.huggingface import HuggingFaceEmbedding

model = HuggingFaceEmbedding(
    model_name="llamaindex/vdr-2b-v1",
    device="cpu",  # "mps" for mac, "cuda" for nvidia GPUs
    trust_remote_code=True,
)

image_embedding = model.get_image_embedding("image.png")
query_embedding = model.get_query_embedding("some query")

via HuggingFace Transformers

from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
from PIL import Image
import torch
import math

# more pixels -> better embeddings -> more VRAM -> slower inference
# From my experience, 768 image patches is the right spot for compute efficient embeddings.
max_pixels = 768 * 28 * 28
min_pixels = 1 * 28 * 28

# Load the embedding model and processor
model = Qwen2VLForConditionalGeneration.from_pretrained(
    'llamaindex/vdr-2b-v1',
    # These are the recommended kwargs for the model, but change them as needed
    attn_implementation="flash_attention_2",
    torch_dtype=torch.bfloat16,
    device_map="cuda:0"
).eval()

processor = AutoProcessor.from_pretrained(
    'llamaindex/vdr-2b-v1',
    min_pixels=min_pixels,
    max_pixels=max_pixels
)

model.padding_side = "left"
processor.tokenizer.padding_side = "left"

document_prompt = "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>What is shown in this image?<|im_end|>\n<|endoftext|>"

query_prompt = "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>Query: %s<|im_end|>\n<|endoftext|>"

Encode queries

def encode_queries(queries: list[str], dimension: int) -> torch.Tensor:
    """
    Encode a list of queries into a tensor of embeddings.

    Args:
        queries: A list of strings, each representing a query.
        dimension: The desired dimension of the output embeddings.

    Returns:
        A tensor of shape (num_queries, dimension) containing the encoded queries.
    """

    dummy_image = Image.new('RGB', (56, 56))
    inputs = processor(
        text=[query_prompt % x for x in queries],
        images=[dummy_image for _ in queries],
        videos=None,
        padding='longest',
        return_tensors='pt'
    ).to('cuda:0')

    cache_position = torch.arange(0, len(queries))
    inputs = model.prepare_inputs_for_generation(
        **inputs, cache_position=cache_position, use_cache=False)

    with torch.no_grad():
        output = self.model(
            **inputs,
            return_dict=True,
            output_hidden_states=True
        )

    embeddings = output.hidden_states[-1][:, -1]
    return torch.nn.functional.normalize(embeddings[:, :dimension], p=2, dim=-1)

Encode documents

def round_by_factor(number: float, factor: int) -> int:
    return round(number / factor) * factor

def ceil_by_factor(number: float, factor: int) -> int:
    return math.ceil(number / factor) * factor

def floor_by_factor(number: float, factor: int) -> int:
    return math.floor(number / factor) * factor

def smart_resize(height: int, width: int) -> tuple[int, int]:
    h_bar = max(28, round_by_factor(height, 28))
    w_bar = max(28, round_by_factor(width, 28))
    if h_bar * w_bar > max_pixels:
        beta = math.sqrt((height * width) / max_pixels)
        h_bar = floor_by_factor(height / beta, 28)
        w_bar = floor_by_factor(width / beta, 28)
    elif h_bar * w_bar < min_pixels:
        beta = math.sqrt(min_pixels / (height * width))
        h_bar = ceil_by_factor(height * beta, 28)
        w_bar = ceil_by_factor(width * beta, 28)
    return w_bar, h_bar

def resize(image: Image.Image):
    new_size = smart_resize(image.height, image.width)
    return image.resize(new_size)

def encode_documents(documents: list[Image.Image], dimension: int):
    """
    Encode a list of images into a tensor of embeddings.

    Args:
        documents: A list of PIL Image objects.
        dimension: The desired dimension of the output embeddings.

    Returns:
        A tensor of shape (num_documents, dimension) containing the encoded images.
    """
    
    inputs = processor(
        text=[document_prompt] * len(documents),
        images=[resize(x) for x in documents],
        videos=None,
        padding='longest',
        return_tensors='pt'
    ).to('cuda:0')

    cache_position = torch.arange(0, len(queries))
    inputs = model.prepare_inputs_for_generation(
        **inputs, cache_position=cache_position, use_cache=False)

    with torch.no_grad():
        output = self.model(
            **inputs,
            return_dict=True,
            output_hidden_states=True
        )
    
    embeddings = output.hidden_states[-1][:, -1]
    return torch.nn.functional.normalize(embeddings[:, :dimension], p=2, dim=-1)

via SentenceTransformers

from sentence_transformers import SentenceTransformer

model = SentenceTransformer(
    model_name_or_path="llamaindex/vdr-2b-v1",
    device="cuda",
    trust_remote_code=True,
    # These are the recommended kwargs for the model, but change them as needed if you don't have CUDA
    model_kwargs={
        "torch_dtype": torch.bfloat16, 
        "device_map": "cuda:0", 
        "attn_implementation": "flash_attention_2"
    },
)

embeddings = model.encode("image.png")

Training

The model is based on MrLight/dse-qwen2-2b-mrl-v1 and it was trained on the new vdr-multilingual-train english subset that consinsists of 100k high quality samples. It was trained for 1 epoch using the DSE approach, with a batch size of 128 and hard-mined negatives.

Results

The model has been evaluated on the Vidore benchmark. All evaluations are performed by calculating NDCG@5 scores using an image resolution that can be represented with maximum 768 tokens.

On the full Vidore benchmark (evaluated with 768 image tokens), both the multilingual and the english-only version performs better than the base model.

	Avg	shiftproject	government	healthcare	energy	ai	docvqa	arxivqa	tatdqa	infovqa	tabfquad
dse-qwen2-2b-mrl-v1	83.6	79.8	95.7	96.9	92	98.2	56.3	85.2	53.9	87.5	90.3
vdr-2b-multi-v1	84.0	82.4	95.5	96.5	91.2	98.5	58.5	84.7	53.6	87.1	92.2
vdr-2b-v1	84.3	83.4	96.9	97.2	92.6	96.8	57.4	85.1	54.1	87.9	91.3

	Avg	shiftproject	government	healthcare	energy	ai
dse-qwen2-2b-mrl-v1 (2560 image tokens)	93.0	82	96	96.4	92.9	97.5
vdr-2b-v1 (768 image tokens)	93.4	83.4	96.9	97.2	92.6	96.8

vdr-2b-v1 matches the performance of the base model on vidore synthetic datasets, while only using 30% of the image tokens (768 vs. 2560). This results in 3x faster inference and much lower VRAM usage.