OpenVLCLIP

An independent Python library that provides an OpenCLIP-inspired inference API for Google's EmbeddingGemma 2 model. It lets Python applications use familiar image and text encoding calls while retaining EmbeddingGemma's own tokenizer, image processing, task prompts, and embedding behavior.

Features

• OpenCLIP-style create_model, create_model_and_transforms, and get_tokenizer entry points.
• Factory discovery with list_models(), list_pretrained(), and create_model_from_pretrained().
• Only EmbeddingGemma 2 is supported; this is not a complete drop-in replacement for OpenCLIP.
• Image and text inference use the official Transformers model and processor; Sentence Transformers is the pinned numerical reference.
• Model API includes forward(), get_logits(), and factory option output_dict=True.
• forward() and get_logits() require explicit logit_scale_mode="cosine" (default: "strict").
• forward() returns normalized features and scale 1.0; get_logits() returns unscaled cosine similarities, not trained CLIP logits or probabilities.
• Native 768-dimensional embeddings plus MRL truncation to 512, 256, or 128 dimensions, with CPU or CUDA device selection.
• Optional text-only or text-and-image model loading; the default factory load remains unchanged.
• A pinned EmbeddingGemma 2 model revision; model weights are downloaded from Hugging Face on first use and are not bundled with this package.
• Bounded-batch image embedding and incremental Japanese/English text-to-image search.
• Optional PyAV-based video-frame indexing and text search.

Installation

Install the package into an active virtual environment. Python 3.12 and later are supported, up to but not including Python 3.15, except Python 3.14.1 (>=3.12,<3.15,!=3.14.1).

pip install openvlclip

Install the optional video dependency for video-frame indexing and search:

pip install "openvlclip[video]"

To pin the exact version for reproducible environments:

pip install openvlclip==0.5.0

The first model or tokenizer load downloads the required files from Hugging Face unless they are already cached. The package pins the model revision 914f7f89142e33e77833254d9c9b90c3cef7303b.

For migration from the v0.4.0 distribution and import namespace, see the v0.5.0 migration guide. openvlclip is a separate distribution; installing it does not uninstall openclip-gemma or provide an openclip_gemma import alias.

Batch image embedding, image search, and video-frame search are documented in the batch embedding guide, image search guide, and video search guide.

Quick Start

This example creates an image in memory, embeds it and a few candidate descriptions, and computes cosine similarities. It requires no external image file.

from PIL import Image, ImageDraw
import torch

from openvlclip import create_model_and_transforms, get_tokenizer

device = "cpu"
model, _, preprocess = create_model_and_transforms(
    "embeddinggemma-2",
    device=device,
)
tokenizer = get_tokenizer("embeddinggemma-2")

image = Image.new("RGB", (256, 256), "navy")
ImageDraw.Draw(image).ellipse((64, 64, 192, 192), fill="red")
image_tensor = preprocess(image).unsqueeze(0).to(device)

descriptions = [
    "a red circle on a blue background",
    "a green triangle on a white background",
]
text_tokens = tokenizer(descriptions).to(device)

with torch.inference_mode():
    image_features = model.encode_image(image_tensor, normalize=True)
    text_features = model.encode_text(text_tokens, normalize=True)
    cosine_similarities = image_features @ text_features.T

print(image_features.shape)       # torch.Size([1, 768])
print(text_features.shape)        # torch.Size([2, 768])
print(cosine_similarities)

The in-memory graphic only demonstrates how to call the API; it is not evidence of the model's image-recognition accuracy. For real use, pass an image relevant to your application.

create_model_and_transforms() returns (model, preprocess_train, preprocess_val). This inference-only package uses the same official preprocessing for both transforms. encode_image() expects the output of the supplied preprocessing function; do not resize or normalize that tensor a second time.

The examples use CPU by default. To request CUDA, set device="cuda" when creating the model and move input tensors to the same device. CUDA inference was tested on the specific GPU listed below; this result does not guarantee support for every NVIDIA GPU or driver combination.

Japanese text

EmbeddingGemma 2 supports multilingual text, including Japanese. Use the returned Gemma token IDs rather than CLIP token IDs or raw strings when calling encode_text():

japanese_tokens = tokenizer(["青い背景にある赤い円"]).to(device)

with torch.inference_mode():
    japanese_features = model.encode_text(japanese_tokens, normalize=True)

print(japanese_features.shape)  # torch.Size([1, 768])

Image and text similarity

The model returns normalized sentence embeddings. You can also request normalization explicitly with normalize=True; normalized image and text features can be compared with a matrix product, as in the Quick Start. A larger cosine similarity indicates a closer match among the supplied candidates, but is not a calibrated probability.

Matryoshka Representation Learning (MRL)

encode_image() and encode_text() accept a keyword-only truncate_dim parameter. Supported values are 768, 512, 256, and 128. The default None keeps the existing 768-dimensional output exactly; lower dimensions take the leading dimensions of the native embedding and L2-normalize the truncated vector. Use the same dimension for image and text features before comparing them:

with torch.inference_mode():
    image_features = model.encode_image(image_tensor, truncate_dim=256)
    text_features = model.encode_text(text_tokens, truncate_dim=256)
    cosine_similarities = image_features @ text_features.T

forward() and get_logits() accept the same truncate_dim when explicitly constructed with logit_scale_mode="cosine". The strict default and the untrained cosine scale behavior are unchanged.

Model loading can omit unused towers with create_model(modalities=...):

from openvlclip import create_model

text_model = create_model(modalities=("text",))
image_text_model = create_model(modalities=("text", "image"))

modalities=None remains the default and preserves the previous full checkpoint load. Text is required because the text transformer is shared; image-only and audio inference are not supported. A text-only model raises an explicit error from encode_image(). Factories that return image transforms require the image modality.

model.capabilities is a read-only mapping with image_embedding, text_embedding, image_text_similarity, patch_embedding, text_similarity_map, mrl, and selective_loading flags. Its values reflect the loaded modalities: a text-only model reports image embedding and image/text similarity as unavailable.

Zero-shot image classification

For a set of candidate labels, encode images and candidate descriptions, then rank each image's candidate scores:

from PIL import Image, ImageDraw
import torch

from openvlclip import create_model_and_transforms, get_tokenizer

device = "cpu"
model, _, preprocess = create_model_and_transforms(
    "embeddinggemma-2",
    device=device,
)
tokenizer = get_tokenizer("embeddinggemma-2", prompt_name="Classification")

images = []
for color in ("red", "green"):
    image = Image.new("RGB", (256, 256), "white")
    ImageDraw.Draw(image).rectangle((48, 48, 208, 208), fill=color)
    images.append(image)

image_batch = torch.stack([preprocess(image) for image in images]).to(device)
candidate_tokens = tokenizer(["a red square", "a green square"]).to(device)

with torch.inference_mode():
    image_features = model.encode_image(image_batch, normalize=True)
    candidate_features = model.encode_text(candidate_tokens, normalize=True)
    scores = image_features @ candidate_features.T

print(scores.argmax(dim=-1))  # Index of the highest-scoring candidate per image

This example illustrates scoring only; its synthetic images do not establish recognition quality. If you apply softmax to candidate scores for display, those values are relative to the candidate set and are not calibrated probabilities. A runnable example with real-image and demo options is available at examples/zero_shot_classification.py.

Task-specific prompts

Task prompts are part of EmbeddingGemma's text encoding behavior. The tokenizer accepts prompt_name; the default is no task prompt. For example:

search_tokenizer = get_tokenizer("embeddinggemma-2", prompt_name="SearchQuery")
classification_tokenizer = get_tokenizer(
    "embeddinggemma-2",
    prompt_name="Classification",
)

Use the prompt appropriate to the role of the text (for example, a search query or a classification label). Prompt names and behavior are defined by the pinned official model/tokenizer configuration; this is not OpenCLIP's text-prompting behavior.

Image search

Image indexing and text search are provided by the openvlclip package. The CLI demo is available from the repository source (examples are not included in the wheel). From a source checkout, install the development environment and run:

uv sync --locked --group dev
uv run python examples/image_search.py index \
  --images ./photos --output ./image-index-256 --batch-size 8 \
  --embedding-dim 256 --device cpu --recursive
uv run python examples/image_search.py search \
  --index ./image-index-256 --query "富士山と青空" --top-k 5 --device cpu

Indexing again updates added, changed, and deleted files. Change --device cpu to --device cuda to request CUDA inference. Lower --batch-size if GPU memory is limited. Search uses EmbeddingGemma's SearchQuery task prompt and reports unscaled cosine similarity scores, not probabilities. Results with equal scores are ordered by relative image path.

The local index is a versioned NumPy .npz archive containing JSON metadata and float32 embeddings; it is read with pickle loading disabled and replaced atomically after a successful update. Image and video index archives are limited to 512 MiB each; split larger collections into multiple indexes. File changes are detected using relative path, size, and modification time in nanoseconds. Touching or rewriting an unchanged image can trigger an unnecessary re-embedding; rewriting a file while preserving both size and timestamp can evade detection. Remove the index and rebuild it if a full refresh is needed. A damaged image or incompatible/corrupt index stops the operation with an error rather than silently skipping data. See the image search design and storage notes.

The Python API is available from openvlclip.batching and openvlclip.image_search:

from openvlclip import create_model_and_transforms
from openvlclip.batching import iter_image_embeddings
from openvlclip.image_search import build_or_update_image_index, search_images

model, _, preprocess = create_model_and_transforms(device="cpu")
for path, embedding in iter_image_embeddings(model, preprocess, ["photos/a.jpg"], batch_size=8):
    print(path, embedding.shape, embedding.device)

build_or_update_image_index(
    model,
    preprocess,
    "photos",
    "image-index-256",
    batch_size=8,
    embedding_dim=256,
)
for result in search_images(model, "image-index-256", "青空の山", top_k=5):
    print(result.path, f"{result.score:.4f}")

The streaming iterator yields CPU tensors in input order. Image-search results contain raw cosine similarity scores, not probabilities. Image indexes use format v2 and store one supported embedding dimension. Existing v1 indexes remain readable at 768 dimensions. An incremental update cannot mix dimensions; use a separate index directory or convert an existing 768-dimensional index with convert_image_index(). Conversion truncates and renormalizes stored vectors without rerunning the model.

Video frame search (optional video extra)

Video indexing and search sample frames and match them to Japanese or English text. The optional PyAV dependency is installed with pip install "openvlclip[video]". From a source checkout:

uv sync --locked --group dev --extra video
uv run python examples/video_search.py index \
  --videos ./videos --output ./video-index --interval 1.0 \
  --batch-size 8 --embedding-dim 256 --device cpu --recursive
uv run python examples/video_search.py search \
  --index ./video-index --query "人がドリルで穴を開けている" \
  --top-k 5 --device cpu

MP4 is the primary target; .mp4, .m4v, .mov, .mkv, .webm, and .avi paths are considered. Actual container and codec support depends on the FFmpeg build bundled with PyAV. The decoder samples the first frame whose presentation timestamp reaches each interval target and records its PTS-derived time; the requested interval is a target, not a promise that a frame exists at that exact time. Video is decoded sequentially, while only a bounded frame batch is sent through the image encoder. Reduce --batch-size if GPU memory is limited.

Search ranks individual sampled frames by cosine similarity. Timestamps are relative to the container playback start, or the video stream start when the container start is unavailable; if neither is present, the first decoded frame defines zero. It does not understand motion, event order, or temporal relationships, and a scene between sampled timestamps can be missed. The local video-index.npz archive stores frame embeddings and timing metadata separately from image indexes, uses pickle-disabled NumPy loading, and is atomically replaced. Video additions, changes, and deletions are reconciled by relative path, file size, and modification time in nanoseconds; preserving both size and timestamp can evade change detection. See the video search design notes for storage, sampling, validation, and decoder limitations.

The API is available in openvlclip.video_search:

from openvlclip import create_model_and_transforms
from openvlclip.video_search import build_or_update_video_index, search_video

model, _, preprocess = create_model_and_transforms(device="cpu")
build_or_update_video_index(
    model, preprocess, "videos", "video-index", interval_seconds=1.0, batch_size=8
)
for match in search_video(model, "video-index", "人がドリルで穴を開けている", top_k=5):
    print(match.video_path, match.timestamp_seconds, match.score)

Video search ranks individual sampled frames. It does not understand motion, event order, or temporal relationships, and scenes between sampled timestamps can be missed. Search loads the full index into RAM and uses brute-force cosine scoring, so memory and latency grow with collection size. Video indexes use format v3 and support the same MRL dimensions. Existing version-2 indexes remain readable at 768 dimensions; version-1 timestamp indexes are still unsupported. Use convert_video_index() to create a separate, smaller-dimensional index without model inference.

OpenCLIP differences and compatibility

openvlclip offers a partial, OpenCLIP-inspired inference API, not an implementation of OpenCLIP. Only the embeddinggemma-2 model is supported.

list_pretrained() reports ("embeddinggemma-2", "official"); official is this package's alias for its pinned Google checkpoint, not an OpenCLIP pretrained tag. forward() in cosine mode returns normalized features and scalar 1.0, not CLIP logits. get_logits() returns unscaled cosine similarities. Neither API reproduces a trained OpenCLIP logit_scale.

For scenario-by-scenario results and migration examples, see the OpenCLIP compatibility report and v0.3.0 migration guide. For MRL index and factory changes, see the v0.4.0 migration guide and MRL benchmark guide.

Migration example

The factory and encoder calls are intentionally familiar, but use the EmbeddingGemma-specific model name, preprocessing transform, and tokenizer:

# EmbeddingGemma-specific model, processor, and retrieval prompt
model, _, preprocess = create_model_and_transforms(
    "embeddinggemma-2",
    pretrained="official",
    device="cpu",
    logit_scale_mode="cosine",
)
tokenizer = get_tokenizer("embeddinggemma-2", prompt_name="SearchQuery")

image_features = model.encode_image(preprocess(image).unsqueeze(0), normalize=True)
text_features = model.encode_text(tokenizer(["a description"]), normalize=True)
similarity = image_features @ text_features.T

This is not a drop-in replacement: an OpenCLIP tokenizer, checkpoint, or arbitrary OpenCLIP model name cannot be substituted. The cosine mode is explicit because its output is not numerically equivalent to trained CLIP logits.

Supported environments and limitations

CUDA verification results

The CUDA integration tests completed with 2 passed, 0 failed, 0 skipped on the GPU listed above. They compared CPU and CUDA results using the same precision and fixed model revision (914f7f89142e33e77833254d9c9b90c3cef7303b). Both modalities returned finite 768-dimensional embeddings.

Latency measurements used three warm-up runs followed by ten measured runs per modality, with CUDA synchronization around each timed call. Model files were already cached, so model-load times do not include downloading. Peak allocated memory is PyTorch's measurement for the model and inference workload, not total system VRAM use. These timings are measurements from one Windows development PC, not performance guarantees.

Additional input and numerical constraints:

• Image tensors must come from the model's returned preprocessing transform and have shape [B, 3, H, W]. Pass RGB PIL images to the transform; it performs the required conversion and processing. Do not resize, rescale, or normalize the resulting tensor again.
• Every image in one tensor batch must have the same processed dimensions so that the tensors can be stacked.
• Tokenizer output is dynamically padded for each batch, rather than fixed to OpenCLIP's usual context length. The pinned model configuration supports up to 8192 tokens; long inputs and batches can use substantial memory. Pass the tokenizer's output directly to encode_text(); do not substitute CLIP token IDs.
• encode_image() and encode_text() return normalized embeddings by default. normalize=True is available to explicitly normalize the returned features. FP16 is not a supported precision; FP32 is the reference, and BF16 has been separately tested on the stated Windows CPU environment.
• pretrained=None is unsupported and raises an error. The library uses the pinned EmbeddingGemma 2 revision; it does not load arbitrary OpenCLIP checkpoints.

Development and tests

Clone the repository and install the development dependencies with uv:

uv sync --locked --group dev
uv run pytest
uv run ruff check .
uv build

The standard CI does not download model weights. Real-model integration tests require the integration extra and access to the pinned Hugging Face model files (or a populated local cache):

uv sync --locked --group dev --extra integration
$env:OPENCLIP_GEMMA_RUN_INTEGRATION = "1"
uv run pytest tests/integration -v

For a CUDA test environment on Windows, create a separate environment so the existing environment is left untouched. The PyTorch and torchvision versions below were used for the RTX 4070 Ti SUPER verification. Install the CUDA wheels from PyTorch's official CUDA 13.0 index; check PyTorch's installation selector for other supported configurations.

uv venv --python 3.12 .venv-cuda
uv pip install --python .venv-cuda\Scripts\python.exe -e ".[test,integration]"
uv pip install --python .venv-cuda\Scripts\python.exe --reinstall --no-deps `
  --index-url https://download.pytorch.org/whl/cu130 `
  torch==2.14.1 torchvision==0.29.1

.venv-cuda\Scripts\python.exe -c "import torch; cuda=torch.cuda.is_available(); print('PyTorch:', torch.__version__, 'CUDA runtime:', torch.version.cuda, 'CUDA available:', cuda, 'BF16 supported:', torch.cuda.is_bf16_supported() if cuda else False)"
$env:OPENCLIP_GEMMA_RUN_INTEGRATION = "1"
.venv-cuda\Scripts\python.exe -m pytest tests/integration/test_cuda_integration.py -v -s

The test skips CUDA checks when CUDA is unavailable; verify the probe reports torch.cuda.is_available() == True before treating a skip as a hardware validation. This CUDA result does not establish Linux real-model inference support. Linux and macOS real-model inference remain unverified.

License and model terms

The OpenVLCLIP library is distributed under the Apache License 2.0. It does not include EmbeddingGemma 2 model weights or OpenCLIP implementation code. OpenCLIP is distributed under the MIT License; EmbeddingGemma 2 weights are governed by the terms and license stated by Google for the pinned model revision on its Hugging Face model card.

The library license does not replace or alter the separate terms that apply to downloaded model weights or third-party dependencies. Review those terms before use.

About the name

The name OpenVLCLIP reflects the goal of providing familiar interfaces across vision and vision-language foundation models. Version 0.5.0 supports EmbeddingGemma 2 only; it does not claim OpenCLIP affiliation or full OpenCLIP compatibility.

This is an independently developed library, not an official extension of OpenCLIP or a Google-maintained package.

The names OpenCLIP and Gemma refer to their respective projects and technologies. No affiliation or endorsement is implied.