Skip to main content


OpenClip is an source implementation of OpenAI's CLIP.

These multi-modal embeddings can be used to embed images or text.

%pip install --upgrade --quiet  langchain-experimental
%pip install --upgrade --quiet  pillow open_clip_torch torch matplotlib

We can the list of available CLIP embedding models and checkpoints:

import open_clip


Below, I test a larger but more performant model based on the table (here):

model_name = "ViT-g-14"
checkpoint = "laion2b_s34b_b88k"

But, you can also opt for a smaller, less performant model:

model_name = "ViT-B-32"
checkpoint = "laion2b_s34b_b79k"

The model model_name,checkpoint are set in

For text, use the same method embed_documents as with other embedding models.

For images, use embed_image and simply pass a list of uris for the images.

import numpy as np
from langchain_experimental.open_clip import OpenCLIPEmbeddings
from PIL import Image

# Image URIs
uri_dog = "/Users/rlm/Desktop/test/dog.jpg"
uri_house = "/Users/rlm/Desktop/test/house.jpg"

# Embe images or text
clip_embd = OpenCLIPEmbeddings(model_name="ViT-g-14", checkpoint="laion2b_s34b_b88k")
img_feat_dog = clip_embd.embed_image([uri_dog])
img_feat_house = clip_embd.embed_image([uri_house])
text_feat_dog = clip_embd.embed_documents(["dog"])
text_feat_house = clip_embd.embed_documents(["house"])
API Reference:OpenCLIPEmbeddings

Sanity Checkโ€‹

Let's reproduce results shown in the OpenClip Colab here.

import os
from collections import OrderedDict

import IPython.display
import matplotlib.pyplot as plt
import skimage

%matplotlib inline
%config InlineBackend.figure_format = 'retina'

descriptions = {
"page": "a page of text about segmentation",
"chelsea": "a facial photo of a tabby cat",
"astronaut": "a portrait of an astronaut with the American flag",
"rocket": "a rocket standing on a launchpad",
"motorcycle_right": "a red motorcycle standing in a garage",
"camera": "a person looking at a camera on a tripod",
"horse": "a black-and-white silhouette of a horse",
"coffee": "a cup of coffee on a saucer",

original_images = []
images = []
image_uris = [] # List to store image URIs
texts = []
plt.figure(figsize=(16, 5))

# Loop to display and prepare images and assemble URIs
for filename in [
for filename in os.listdir(skimage.data_dir)
if filename.endswith(".png") or filename.endswith(".jpg")
name = os.path.splitext(filename)[0]
if name not in descriptions:

image_path = os.path.join(skimage.data_dir, filename)
image ="RGB")

plt.subplot(2, 4, len(images) + 1)

images.append(image) # Origional code does preprocessing here
image_uris.append(image_path) # Add the image URI to the list